SELF-PERCEIVED JOB INSECURITY AND THE DEMAND FOR MEDICAL REHABILITATION: DOES FEAR OF UNEMPLOYMENT REDUCE HEALTH CARE UTILIZATION?

Authors


ABSTRACT

An inverse relationship between job insecurity and sickness absence has been established in the literature, which is explained by employees avoiding to send signals of both poor health and uncooperative behavior towards the employer. In this paper, we focus on whether the same mechanism applies to the demand for medical rehabilitation measures. This question has recently gained much interest in the context of the current public debate on presenteeism. Using county-level unemployment rates as instrument for the employees' fear of job loss on the individual level, we find that an increase in subjective job insecurity substantially decreases the probability of participating in medical rehabilitation. Copyright © 2013 John Wiley & Sons, Ltd.

1 INTRODUCTION

Since the start of the economic crisis in the fall of 2008, the fear of unemployment has become much more prevalent among employees in the USA and most European countries (Curtin, 2010; EC, 2011). This development is a matter of concern because subjective job insecurity has been shown to significantly deteriorate individual well-being (Knabe and Rätzel, 2011; Geishecker, 2010). Moreover, adverse effects, such as lowered wages and psychological problems, have been documented; see for example, Blanchflower (1991), Blanchflower and Shadforth (2009), Campbell et al. (2007), and Dooley et al. (1987).

Another related question is whether perceived job insecurity also affects the demand for health care services. This question has recently gained much interest in the context of the current public debate on ‘presenteeism’ (CBC, 2010; RC, 2009; Welt, 2010). Because health care utilization requires time off the job, presenteeism may not only mean going to work when sick but also forgoing proper medical treatment. Despite its relevance to the public debate, to the best of our knowledge, this paper is the first one to empirically address the link between subjective job insecurity and the demand for health care services. In doing so, it contributes to the question raised by Ruhm (2003) whether health investment declines when local unemployment rates rise.

In particular, our analysis tests the hypothesis that fear of job loss reduces participation in medical rehabilitation. In line with the notion of presenteeism, the argument is that employees who are afraid of losing their jobs take less rehabilitation leaves in order to avoid sending signals of poor health. Moreover, they may want to avert falling in disgrace of the employer who in most European countries and in certain regions of the USA is obliged to pay sickness benefits.1 Several empirical analyses support this reasoning by establishing a causal link between job insecurity and sickness absence, for example, Ichino and Riphahn (2005), Arai and Thoursie (2005), Askildsen et al. (2005), Fahr and Frick (2007), Bradley et al. (2007), Olsson (2009), Khan and Rehnberg (2009), and Jacob (2010).

We focus on medical rehabilitation because, among all health care services, its consumption implies the strongest unfavorable signals towards the employer. Compared with doctor visits, for instance, it typically involves more sick leave. Even worse, the unique feature of rehabilitation leaves is that the employer may misinterpret them as shirking work because medical rehabilitation often addresses non-acute or invisible health problems. Moreover, the scientific literature has not yet satisfactorily settled whether it is effective (in the long term).2 It is probably therefore that the true need for rehabilitation measures is often called into question in the (German) public (Raspe, 2009).

The empirical analysis is based on individual-level data for the years 2003, 2004, and 2006 from the German Socioeconomic Panel (SOEP). The survey is well-suited for our analysis because it includes information on rehabilitation participation and self-assessed job insecurity, along with numerous individual characteristics. In order to tackle the potential endogeneity of self-assessed job insecurity, we employ a nonlinear instrumental variable estimator, where the unemployment rate at the county level serves as the instrument for job insecurity. We find a pronounced negative and significant effect of job insecurity on participation in medical rehabilitation.

The remainder of the paper is organized as follows. Section 2 discusses the institutional framework of medical rehabilitation in Germany and introduces the data. Section 3 discusses the econometric approach. Section 4 reports the empirical results, and Section 5 concludes.

2 INSTITUTIONAL SETTING AND DATA

Medical rehabilitation covers measures to reduce the physical, psychological, and social consequences of an illness, an accident, or disability in order to regain self-sufficiency, allowing for a life as close to normality as possible. Rehabilitation covers various health problems ranging from cancer, stroke, etc. to non-acute conditions such as drug addiction or exhaustion. In Germany, the majority (34%) of medical rehabilitation treatments address musculoskeletal disorders. About 16%, 12%, and 11% of the medical rehabilitation activities are related to cardiovascular diseases, cancer, and psychological disorders, respectively (Augurzky et al., 2009). The German social pension fund pays for the cost if rehabilitation focuses on the recovery of the patient's ability to work.3 If a particular case does not fall into the jurisdiction of the social pension fund or any other reimbursement authority, patients can claim the reimbursement against their health insurer.4 In essence, this means that the vast majority of individuals in Germany have insurance coverage for medical rehabilitation.

In any case, patients have to apply for rehabilitation measures. Subsequently, the corresponding reimbursement authority must decide on the application on purely medical grounds and – if rehabilitation participation is approved – has to assign a rehabilitation center or another medical facility within 5 weeks. It is important to note that employers have no right to object against any rehabilitation participation by their personnel. Moreover, employees are entitled to continued pay during rehabilitation participation, although many employers do not receive any compensation.5 The duration of rehabilitation is generally 3 weeks depending on medical aspects (DRV, 2008; Augurzky et al., 2009). The duration of treatment may be extended to a maximum length of 7 weeks. However, continued payment of wages by the employer stops after the first 6 weeks. From this week on, the health insurance pays the sickness benefits.

Medical rehabilitation can be pursued on an inpatient or outpatient basis. As far as the medical treatment is concerned, there are no substantial differences between these options. In particular, the specific medical measure does not indicate whether individuals will be assigned to inpatient or outpatient rehabilitation, and treatment is typically full-time in either case. Inpatients have to co-pay €10 per day up to a maximum of 42 days, provided they are in an economic position to do so.6 For outpatient rehabilitation, there is no co-payment schedule. Roughly 85% of all rehabilitation measures are inpatient procedures (Augurzky et al., 2009).7

The analysis is based on data from the German SOEP, a large longitudinal household survey that started in 1984 (Haisken-DeNew and Frick, 2005). The SOEP includes a wide range of information at the individual and the household level such as working and living conditions, as well as variables on individual health status and health care utilization. We focus on the waves of 2003, 2004, and 2006 as information on the outcome variable is available only for these years.8 The dependent variable is a binary indicator taking the value of 1 if the respondent was in rehabilitation at least once during a particular calendar year and 0 otherwise.

In the SOEP, individuals are asked whether they are (i) not concerned at all, (ii) somewhat concerned, or (iii) very concerned about their job security. We use the corresponding answer as the key explanatory variable in our empirical analysis. Hence, our measure of fear of unemployment is measured on an ordinal scale. The survey includes this question annually.9

Most relevant to our econometric approach, the data contain the annual average unemployment rate that corresponds to the county of residence. Currently, Germany is sectioned into 413 counties (‘Kreise’), which represent key territorial units for public planning and administration. Average county population and average county territory is 198,000 inhabitants and 865 km2, respectively.10 The county unemployment rate, taking values that range from 3.4% to 27.5%, is used as the instrumental variable for the fear of unemployment, see Section 3. As the variable fear of unemployment is linked to the time of the interview, rather than taking the annual county unemployment rate, we use a weighted average over the previous and the current year. In detail, we calculate math formula, where Uct is the annual average unemployment rate in individual i's home county c, whereas t and M denote the relevant year and the month of interview, respectively.

As control variables, we use socioeconomic characteristics, such as sex, age, years of education, a dummy indicating being born abroad, household size, and an indicator for living together with a partner. We also include the number of children younger than 18 years and marital status (married and non-married). This is done because dismissal protection is especially strict for married people and those with dependent children. Hence, both are potential determinants of individual fear of job loss. One may also think of direct effects on rehabilitation participation, for example, individuals with small children may hesitate to leave the family for several weeks.

We also control for the working environment in order to account for individual differences in dismissal protection. The working environment may also directly affect the demand for medical rehabilitation. First, we use a set of dummy variables capturing firm size, that is, (i) up to five, (ii) more than five, and (iii) more than 2000 employees. Here, small firms serve as the reference category, as small firms are exempted from stringent dismissal protection regulation. As both medium-sized and large companies are subject to equivalent dismissal protection regulations, the latter exhibits particularly low dismissal rates (Bachmann and David, 2010). Other working environment variables closely related to individual job insecurity are firm tenure and a dummy indicating a temporary contract. Besides these, we control for holding a side job as well as for marginal employment (‘mini-job’ or ‘midi-job’), which is often less stable than ordinary employment. We also include a set of dummies capturing occupation, that is, (i) unskilled blue-collar, (ii) skilled blue-collar, (iii) low-skilled white-collar, and (iv) high-skilled white-collar, where the first serves as reference. Personal gross labor income, measured in €1000 per month, also enters the empirical model as a control.

Individual health certainly represents a key determinant of rehabilitation participation, yet it is potentially linked to subjective job insecurity, as well. Hence, we control for health by a set of dummies indicating self-assessed health, ranging from (i) bad to (v) very good. As another health related variable, the officially approved degree of disability enters the model. This variable matters for institutional reasons, as particularly stringent dismissal protection regulations apply for severely handicapped employees. In order to avoid potential bias resulting from reversed causality, health variables and income enter the analysis in terms of 1-year lags.11 In addition, year and state dummy variables are considered. The latter control for regional differences in accepting applications for medical rehabilitation.12

An alternative model specification includes additional variables, controlling for the availability of rehabilitation centers and regional characteristics. In particular, the county mortality, the county age distribution, and indicators for the grade of rurality are used. Moreover, in the extended model specification, we control for the health insurance status. We distinguish public and private health insurances, where the former serves as reference. Legally employed are always insured, hence, we do not consider a category ‘not insured’. We do not include any price variable because in Germany, almost everyone has health insurance coverage for rehabilitation. As robustness checks, we tested whether the results of the basic model change if the set of control variables is varied. In particular, we included indicators for self-perceived risk aversion as additional regressors and replaced self-assessed health by alternative health proxy variables, respectively. Moreover, in a third variation to the basic model, we excluded the dummy variable indicating a temporary work contract in order to illustrate that it does not take away considerable parts of the effect of job insecurity on rehabilitation demand.13 As shown in Table A1 in the Appendix, results are robust to these alternative model specifications.14

Naturally, we only consider employed individuals in the empirical analysis. We further exclude conscripts, the self-employed, and public servants because these groups cannot be laid off and, thus, are likely to behave differently. The latter have a special status in being protected against dismissal by law. Although public sector employees may possibly be concerned about losing their jobs, we also exclude them from the empirical analysis as, in general, they are hardly ever laid off.15 This is most likely due to enhanced dismissal protection in the public sector, that is, public sector employees with a certain employment duration almost acquire public servant status. In fact, according to our data, they are by far less concerned about their job security than private sector employees. Finally, we exclude individuals that moved into another federal state between the years 2003 and 2006 from the estimation sample in order to reduce the problem of self-selection into areas with certain labor market conditions.16

After excluding these groups of individuals17 and individuals with missing information, the sample for 2003 consists of 6559 individual-level observations. For the year 2004 and 2006, the corresponding numbers are 5585 and 5916 observations, respectively. In 2003, 269 individuals were in rehabilitation. In 2004 and 2006, the numbers of rehabilitation participants were 176 and 192. Regarding the key explanatory variable in 2003, we observe that 45% of the individuals were somewhat and 18% were very concerned about their job security. In 2004, these shares amounted to 46% and 21%, and in 2006, to 45% and 18%. The shares of rehabilitation participants among both employees who were not concerned and employees who were somewhat concerned about their job security amounted to 3.3% and 3.4% (Table 1). In contrast, for employees who were very concerned about their job security, the corresponding figure was 4.1%. These descriptive figures seem to argue against the hypothesis of job worries reducing the demand for medical rehabilitation.

Table 1. Rehab participation by concerns about job security
 Concerns about job security
Rehab participationNot concernedSomewhat concernedVery concerned
  1. Statistics for the sample used for estimating the basic model (18,060 obs.); compare with Section 4.

  2. Source: Own calculations.

No6216(96.7%)7918(96.6%)3289(95.9%)
Yes215(3.3%)280(3.4%)142(4.1%)

The average sample age is 42 years. Almost 50% of the individuals are women. White-collar workers comprise 60%. The average sample individual has 12 years of education and earns a monthly gross wage of 2200. Most (90%) observations have at least a satisfactory health status. Only 13% live in rural areas, whereas 35% live in urban areas with a high population density. For comprehensive descriptive statistics, see Table 2.

Table 2. Descriptive statistics
 MeanS.D.Min.Max.
  1. Statistics for the sample used for estimating the basic model (18 060 obs.); compare with Section 4. For the controls, no statistics for reference categories reported.

  2. Source: Own calculations.

Dependent variable: Rehabilitation participation0.0350.18401
Endogenous regressor: Concerned about job security (Ai)
Not at all0.3570.47901
Somewhat0.4540.49801
Very0.1900.39201
Instrument: County unemployment rate0.1090.0490.0340.275
Controls:
Age (years)41.58110.9411882
Male0.5330.49901
Immigrant0.1330.34001
Years of education12.1412.467718
Married0.6600.47401
Living with partner0.7570.42901
Household size2.9561.263113
Number of kids younger than 18 years0.6840.95008
Personal gross income (€1000 per month, lag)2.2131.749036
Occupation
Blue-collar skilled0.1910.39301
White-collar low-skilled0.4100.49201
White-collar high-skilled0.1820.38601
Tenure (years)10.2539.327049.8
Mini-job0.0740.26201
Midi-job0.0440.20601
Temporary work contract0.0980.29701
Side job0.0570.23101
Unemployed (lag)0.0620.24201
Firm size
Medium (6–2000 employees)0.4850.50001
Large (more than 2000 employees)0.4280.49501
Self-assessed health (lag)
Poor0.0830.27501
Satisfactory0.3090.46201
Good0.4960.50001
Very good0.1020.30301
Degree of occupational disability (lag)0.0230.10901
Year 20040.3090.46201
Year 20060.3280.46901
Private health insurance0.0590.23501
County mortality (deaths per 1000 inhabitants and year)10.1221.4586.36114.943
County age distribution (years)
Between 20 and 500.4240.0240.3560.527
Between 50 and 700.2500.0180.1920.329
Above 700.1250.0150.0830.181
Type of region
Agglomeration0.1950.39601
Urban area (high population density)0.1540.36101
Urban area (medium population density, big cities)0.0630.24301
Urban area (medium population density, no big cities)0.1520.35901
Rural area (high population density)0.0910.28701
Rural area (medium population density)0.0390.19401

3 THE ECONOMETRIC APPROACH

We set up a regression model explaining the intention to participate in rehabilitation math formula by the individual fear of job loss math formula and a vector of control variables Xi:

display math(1)

Here, subscript i indicates the individual, εRi represents a random error term, whereas γ and βR are coefficients subject to estimation. The focus is on γ, measuring the effect of job insecurity. The key explanatory variable math formula may suffer from endogeneity due to unobserved heterogeneity. For instance, highly motivated individuals might assess the probability of losing their jobs more optimistically and, at the same time, might demand less rehabilitation services because they feel indispensable at the job. Another example is company-specific factors such as management style, which may influence job insecurity and the demand for medical rehabilitation. Moreover, reversed causality is a potential issue as rehabilitation participation might affect employment perspectives.

In order to tackle this potential endogeneity problem, we pursue an instrumental variable approach (see Angrist and Krueger 2001), for an overview). That is, we augment Equation ((1) with the equation

display math(2)

explaining the endogenous regressor, where Zi is the instrumental variable. For the errors εAi and εRi, we assume joint normality with error correlation ρ. Any value of ρ deviating from zero renders math formula an endogenous regressor in Equation (1). The parameters βA and α denote coefficients subject to estimation.

The reasoning behind county unemployment representing an appropriate instrument is that individuals may estimate their individual baseline risk of job loss on the basis of regional unemployment rates. Yet, as meso-level variables, regional labor market indicators are not affected by individual behavior, individual perception, or unobserved company-specific factors, that is, they are exogenous in econometric terms. Moreover, it is unlikely that regional unemployment exerts a direct effect on the individuals' demand for rehabilitation that does not operate through the endogenous explanatory variable.18

Yet, the impact of regional labor market conditions on self-assessed job insecurity is likely to be heterogeneous across individuals. For this, individual character traits as well as institutional and job-specific factors may play a role. We partly account for these factors by including job-related variables. Moreover, we exclude some groups of individuals from the econometric analysis, for example, public servants, who, for legal reasons, cannot be laid off regardless of how bad the economy is going.

The strength of the instrumental variable may be limited if many employees commute across county boarders and the unemployment rate in the county of the workplace predominantly matters for the perception of the security of their jobs. Although, in fact, commuting between work and home over the county boarder applies to roughly 40% of all German employees (Haas and Hamann, 2008), we argue that the unemployment rate in the county of residence is just as relevant as the respective rate in the county of the workplace because cross-county commuters may take best note of dismissals or employment recruitment in their social environment, which is usually in the county of residence.19

Neither math formula (the intention to participate in rehabilitation) nor math formula (the actual level of fear) is directly observed in the data. Rather, we observe the binary variable Ri, indicating actual rehabilitation participation, and the ordinal measure for job insecurity Ai. In order to account for both math formula being potentially endogenous as well as being unobserved, we substitute math formula by (2) in the structural Equation (1) and estimate the resulting reduced form equation

display math(3)

jointly with the instrumental Equation (2) using bivariate ordered probit.20,21 Note that Var(εRi + γεAi) = (1 + 2γρ + γ2) holds for the reduced form error. Because probit estimation implicitly rescales the error, in Equation (3), all estimated coefficients are re-scaled by the factor math formula. Hence, besides estimates for βA and α, the estimation procedure yields estimates for

display math(4)
display math(5)
display math(6)

where the latter is the covariance between the re-scaled reduced form error and εAi. Solving Equations (4)(6) and inserting the estimates for βA, α, math formula, math formula, and math formula yield maximum likelihood estimates for γ, βR, and ρ. Obviously, identification of the structural parameters rests on both valid exclusion restrictions and joint estimation; see Sajaia () for a more detailed description of the estimation procedure. In order to account for the unemployment rate being measured at the county level, which is likely to result in underestimated standard errors (Moulton, 1986), we compute robust standard errors by clustering the data by county.22

Although the suggested approach directly allows for determining marginal effects for the controls, a problem remains with the estimated marginal effect for math formula. The problem is that the size of its fitted values math formula has no economic meaning because the coefficient estimates are implicitly scaled by an unknown standard deviation. In other words, we do not know the scale on which to measure a marginal increase in fear of unemployment. Therefore, rather than considering a one-unit increase in math formula, we use the estimated ‘inter-threshold-range’ (ITR) as unit of reference. This measure for the effect size is robust to the scaling of math formula. In more detail, ITR is the distance between the two threshold values – estimated by applying ordered probit to Equation (2) – that separate the three observed categories of fear. That is, 1 ITR represents the increase (respectively, the decrease) in math formula that is required to push any individual from the category ‘somewhat concerned’ to the category ‘very concerned’ (respectively, ‘not concerned at all’). This, however, is a large, non-marginal change.23 Hence, rather than considering a full ITR change in math formula, we consider one-tenth of it when calculating marginal effects on the likelihood of rehabilitation participation.24 The reported effect represents the sample mean of individual effects. The corresponding confidence interval is bootstrapped, where clustering by county is taken into account.

4 RESULTS

In this section, we discuss results for our basic specification and the extended specification, which includes further control variables. Prior to that, we validate our identification strategy by comparing reduced form results for private sector employees and public servants.

4.1 Validity of the instrument

Identification of the effect rests on the assumption that county unemployment is, conditional on the included covariates, linked to rehabilitation participation exclusively through individual fear of job loss. As we only have one instrumental variable at hand, this assumption is not directly testable. Fortunately, we can indirectly assess whether county unemployment operates through other channels than the fear of unemployment. Because public servants are strictly protected against dismissal and, in turn, will by no means be concerned about becoming unemployment, county unemployment is uncorrelated with rehabilitation participation for this group of individuals if the instrument is valid. This does not imply that for public servants, job-related worries are immaterial for the decision whether or not to participate in medical rehabilitation. For instance, public servants may be concerned about a less favorable assessment by their superior and reduced promotion prospects if they are off the job for a substantial period of time. However, for public servants – unlike private sector employees – such concerns are not related to the business cycle and regional unemployment.

In order to indirectly test the validity of the instrument, we estimate the reduced form model (3) for the actual estimation sample consisting of private sector employees and compare the estimated coefficient of county unemployment to its counterpart estimated for public servants. Results for both regressions are presented in Table 3. The coefficient of county unemployment is significant and negative for private sector employees (left part of Table 3). In contrast, for public servants, it is positive and insignificant (p = 0.434). Although the sample size for the latter is relatively small, we still can clearly reject the null of equal coefficients. In fact, regional unemployment is of significantly greater relevance for the participation decision of private sector employees.25 This is indication for the instrument being valid.

Table 3. Reduced form: rehab participation of private sector employees and public servants
 Private sector employees (estimation sample)Public servants
 CoefficientS.E.CoefficientS.E.
  • ***

    significant at 1%;

  • **

    significant at 5%;

  • *

    significant at 10%; standard errors clustered by county; estimated by univariate probit.

  • Source: Own calculations.

County unemployment−3.467***0.8491.5501.982
Age0.010***0.0020.0060.012
Male−0.0050.0430.0500.131
Immigrant−0.207***0.073
Years of education−0.0100.010−0.0090.023
Married−0.0460.058−0.1870.206
Living with partner0.0010.0660.2620.219
Household size−0.0330.0280.1010.062
Number of kids younger than 18 years0.0480.032−0.0080.087
Personal income (lag)0.0120.016−0.0500.060
Occupation
Blue-collar skilled−0.0710.061
White-collar low-skilled−0.106*0.056
White-collar high-skilled−0.165*0.085
Tenure−0.0010.0020.0080.009
Mini-job−0.321***0.093
Midi-job−0.0110.095−0.3350.472
Temporary work contract−0.0050.072
Side job0.0080.0870.2980.230
Unemployed (lag)−0.0570.107−0.1480.430
Firm size
Medium0.1270.082−0.8040.631
Large0.218***0.083−0.8160.629
Self-assessed health (lag)
Poor−0.1440.116−0.3070.364
Satisfactory−0.431***0.113−0.4410.344
Good−0.662***0.113−0.702**0.341
Very good−0.925***0.144−0.740*0.388
Degree of occupational disability (lag)0.712***0.121−0.3890.431
Year 2004−0.090**0.045−0.0580.129
Year 2006−0.081*0.044−0.0140.147
Constant−1.099***0.271−1.0771.021
State dummiesYes***Yes
Number of observations18 0601710

4.2 The effect of job insecurity on the demand for medical rehabilitation

The left part of Table 4 illustrates the estimation results of primary interest, that is, results for the basic specification. Concerning the instrumental equation (second part of the table), the county unemployment rate exhibits the expected positive sign. The test on instrument relevance turns out to be highly significant. The relevant test statistic exceeds by far the threshold of 10, which is suggested by Staiger and Stock (1997).26 Thus, we are not concerned about the instrument being weak.

Table 4. Rehab participation: estimated coefficients
 Basic modelExtended model
 CoefficientS.E.CoefficientS.E.
  • ***

    significant at 1%;

  • **

    significant at 5%;

  • *

    significant at 10%. Standard errors clustered by county.

  • Source: Own calculations.

Structural equation: rehabilitation participation
Fear of unemployment−0.782***0.110−0.776***0.106
Age0.0020.0030.0020.003
Male0.0390.0360.0390.036
Immigrant0.0510.0750.0680.075
Years of education−0.014*0.008−0.0120.008
Married0.0060.045−0.0050.045
Living with partner0.0250.0500.0310.051
Household size−0.0190.020−0.0250.020
Number of kids younger than 18 years0.044*0.0230.048**0.024
Personal income (lag)0.0030.0130.0090.013
Occupation
Blue-collar skilled−0.0380.047−0.0420.047
White-collar low-skilled−0.207***0.044−0.198***0.044
White-collar high-skilled−0.332***0.067−0.322***0.067
Tenure−0.006***0.002−0.006***0.002
Mini-job−0.567***0.068−0.563***0.068
Midi-job−0.163**0.074−0.158**0.074
Temporary work contract0.141***0.0530.141***0.052
Side job−0.0650.060−0.0620.060
Unemployed (lag)−0.0010.0750.0060.075
Firm size
Medium0.174***0.0610.170***0.061
Large0.215***0.0650.220***0.066
Year 20040.0330.0430.0630.046
Year 2006−0.0470.0410.0060.056
Self-assessed health (lag)
Poor−0.0360.109−0.0380.108
Satisfactory−0.336***0.111−0.340**0.110
Good−0.583***0.120−0.588***0.118
Very good−0.913***0.138−0.915***0.136
Degree of occupational disability (lag)0.2600.1720.2650.171
Private health insurance−0.0770.068
County mortality0.0430.027
County age distribution (years)
Between 20 and 50−2.054*1.216
Between 50 and 70−0.3711.605
Above 70−4.396*2.534
Type of region
Agglomeration0.0910.077
Urban area (high population density)0.0760.064
Urban area (medium population density, big cities)0.136*0.079
Urban area (medium population density, no big cities)0.139**0.064
Rural area (high population density)0.1060.080
Rural area (medium population density)0.0670.085
Constant−0.669***0.249−0.4070.723
Instrumental equation: fear of unemployment
County unemployment2.696***0.6833.068***0.684
Age−0.005***0.001−0.005***0.001
Male0.053**0.0250.053**0.025
Immigrant0.226***0.0440.239***0.043
Years of education−0.010*0.006−0.0070.006
Married0.0440.0350.0370.035
Living with partner0.0310.0370.0340.037
Household size0.0020.014−0.0020.014
Number of kids younger than 18 years0.0200.0190.0210.019
Personal income (lag)−0.0060.009−0.0010.010
Occupation
Blue-collar skilled0.0060.0340.0050.034
White-collar low-skilled −0.183***0.033−0.175***0.033 
White-collar high-skilled−0.296***0.045−0.289***0.045
Tenure−0.006***0.001−0.007***0.001
Mini-job−0.477***0.048−0.475***0.049
Midi-job−0.200***0.042−0.199***0.042
Temporary work contract0.184***0.0370.186***0.037
Side job−0.089**0.038−0.090**0.038
Unemployed (lag)0.0440.0370.0480.037
Firm size
Medium0.124***0.0390.123***0.039
Large0.105***0.0410.111***0.041
Year 20040.113***0.0200.128***0.025
Year 20060.0030.0260.0330.043
Self-assessed health (lag)
Poor0.0650.1040.0630.103
Satisfactory−0.0950.101−0.1000.101
Good−0.231**0.102−0.236**0.102
Very good−0.449***0.104−0.450***0.103
Degree of occupational disability (lag)−0.220**0.094−0.218**0.094
Private health insurance–0.0600.051
County mortality0.0230.025
County age distribution (years)
Between 20 and 50−1.997**0.965
Between 50 and 700.3361.296
Above 70−3.5522.234
Type of region
Agglomeration0.0720.076
Urban area (high population density)0.0440.065
Urban area (medium population density, cities)0.173**0.077
Urban area (medium population density, no big cities)0.0410.069
Rural area (high population density)0.1280.079
Rural area (medium population density)0.0150.089
State dummiesYes***Yes***
ρ0.794***0.1070.789***0.103
Threshold values ordered probit (instrumental equation)−0.2280.185−1.0080.707
1.096***0.1850.3200.708
Joint significance (x2-statistic)55234325
Number of observations18 06018 032

With respect to the controls, we find that age, tenure, and being disabled exhibit the expected negative effect on fear of being laid off. We do not observe a significant effect either of being married or of the number of underage dependents. This result might be explained by people who bear family responsibilities, on the one hand, being genuinely more concerned about unemployment, but on the other hand, being particularly well protected by the relevant regulations. Somewhat surprisingly, staff employed at small firms seem to be less concerned about job loss. Moreover, the marginally employed and those who hold a side job are less worried about becoming redundant. As expected, white-collar and highly qualified workers feel more secure about their jobs than blue-collar and low-qualified workers. Better self-assessed health goes along with less worries about job loss. Moreover, gender, region, and immigrant status also matter. The finding of men being more concerned may reflect their traditional gender role as the family's major bread-winner. A significantly higher level of subjective job insecurity is found in eastern federal states, which might be explained by the economy being generally weaker in this part of the country. The estimated error correlation math formula is positive and deviates significantly from zero. Hence, endogeneity of the key explanatory variable is a relevant issue.

Considering the equation of primary interest (first part of Table 4), as expected, self-assessed health turns out to be a major and highly significant determinant of rehabilitation participation (p-value for the test on joint significance: 0.000). Moreover, rehabilitation utilization is significantly more prevalent among employees of large firms. This points at intra-firm (time) constraints exerting a substantial impact on the decision to take a rehabilitation-related sick leave. The result of highly skilled employees being less likely to participate in rehabilitation points in the same direction. Interestingly, although we control for self-assessed health and several other characteristics, estimated state effects indicate that employees living in East Germany are significantly more likely to participate in rehabilitation compared with those living in the western part of the country. In terms of marginal effects, the East–West differential amounts to roughly 6 percentage points.

Most importantly, the coefficient for fear of unemployment is negative and highly significant in statistical terms. This is evidence in favor of our hypothesis. Employees are apparently less reluctant to participate in rehabilitation when there is no need to secure their position. In quantitative terms, the mean effect of a change in fear of unemployment of one-tenth of the estimated ITR amounts to − 1.25.27 This means that on average, the probability of participating in rehabilitation would decrease by 1.25 percentage points if the self-assessed probability of job loss increased by one-tenth of the ITR. An increase in fear of this magnitude might, for instance, be attributed to a change in the regional unemployment rate of approximately 5 percentage points.28 Given the small relative frequency of rehabilitation participation, 1.25 percentage points represent a proportional decrease of nearly 35%.

The extended model gives support to the negative effect of job insecurity on medical rehabilitation, see the right part of Table 4. The coefficient of fear of unemployment is highly significant and, in quantitative terms, nearly identical to the coefficient estimated for the basic model. The additional covariates are jointly and – with few exceptions – individually insignificant in the structural equation. The local availability of rehabilitation centers and clinics seems not to represent a relevant factor for the decision to participate in rehabilitation. Neither does the health insurance status seem to matter for participating in medical rehabilitation. In contrast, the additional controls are jointly significant (p = 0.02) in the instrumental equation. They indicate that individuals living in counties with a high share of inhabitants aged between 20 and 50 years (that is, the region is characterized by a relatively large workforce) are less concerned about their job security. This is probably due to alternative job opportunities. In line with this, fear of unemployment varies across regions of different degrees of rurality.

4.3 Marginal effects by health status

The question remains whether less participation results from forgoing a treatment that is highly beneficial in medical terms or from forgoing a less important one. The available data limit the opportunities for directly addressing this issue because we do not observe in what kind of rehabilitation measures employees participate and, therefore, cannot observe what kind of treatments are forgone because of job insecurity. Nevertheless, we try to shed some light on this question by comparing the strength of the effect across individuals of different health status. One may argue that individuals in poor health are, on average, in greater need of health care. And, in turn, if the fear of unemployment did first of all reduce the consumption of rehabilitation services of questionable or limited benefit, the effect should be more prominent for healthy individuals. One may, however, object that this argument does not apply to medical rehabilitation because it addresses many (chronic) diseases that are more effectively treated if they are not very advanced as compared with final stages of the illness. Consequently, individuals in generally good health may also substantially benefit from rehabilitation utilization.29 In this case, the finding of differential effects between groups of different health status would be inconclusive about whether job insecurity primarily reduces the consumption of rehab measures of questionable medical benefit.

Table 5 displays estimated marginal effects stratified by the 1-period lagged self-assessed health status.30 We consider both estimates derived from the pooled model discussed previously, that is, the coefficients do not vary across health states, as well as estimates derived from re-estimating the model separately for individuals of different health states. The point estimates – quite contrary to the previous argument of people in poor health being in greater need of health care – seem to argue in favor of marginal effects decreasing with better health.31 However, for both, pooled and stratified estimations, deviations in mean marginal effects are jointly insignificant at the 0.05 level. When only comparing marginal effects for the extreme health states, the pooled model yields a significant (p = 0.021) differential, but stratified estimation does not (p = 0.112). Thus, the results certainly do not indicate stronger effects for healthy individuals. In contrast, they can be regarded as weak evidence for the effect of job insecurity diminishing with better health.

Table 5. Mean marginal effect of fear of job loss over self-assessed health
 Self-assessed health status (lag)
 BadPoorSatisfactoryGoodVery Good
 Marg. eff.Marg. eff.Marg. eff.Marg. eff.Marg. eff.
  • ***

    significant at 1%;

  • **

    significant at 5%;

  • *

    significant at 10%; marginal effects measured as percentage points; standard errors in parentheses;

  • a

    categories ‘bad’ and ‘poor’ consolidated because of insufficient data for the former.

  • b

    categories ‘good’ and ‘very good’ consolidated because of insufficient variation in the dependent variable for the latter.

  • Source: Own calculations.

Pooled estimation−3.327**−2.588**−1.579**−0.934**−0.479**
(1.399)(1.065)(0.626)(0.381)(0.209)
Stratified estimation−3.424*a−2.399**−0.558b
(1.747)(0.906)(0.458)

5 CONCLUSIONS

This paper contributes to the research on the effects of job insecurity on sickness absence by adding another aspect to the existing literature. We analyze whether the pattern of strategic behavior of employees, that is, avoiding to send signals of poor health and uncooperative behavior towards the employer by going to work, also applies to the utilization of health care services. This is reasonable because consuming health services typically implies absence from work.

With a special focus on the demand for medical rehabilitation, our empirical results support this line of argument. We identify a negative effect of subjective job insecurity on the likelihood to participate in medical rehabilitation. In quantitative terms, the probability of participating in rehabilitation would decrease by 1.25 percentage points if the self-assessed probability of job loss increased by one-tenth of the estimated inter-threshold range. An increase in fear of this magnitude might, for instance, be attributed to a change in the regional unemployment rate of roughly 5 percentage points. Considering the rather low average rate of rehabilitation utilization, this represents a proportional decrease of nearly 35%. This is evidence for individuals decreasing health investment in bad economic times as hypothesized by Ruhm (2003).

Individual behavior of forgoing medical treatment if the job is insecure may be interpreted quite differently, depending on whether a necessary or a treatment of limited medical value is forgone. Although in the former case, presenteeism is practised at its worst, most likely implying adverse health effects, in the latter, excess utilization of rehabilitation measures, that is, moral hazard, is reduced. Although little is known about the extent of moral hazard in the utilization of medical rehabilitation in Germany, apart from a likely reduction through a doubling in daily co-payments in 1997 (Ziebarth, 2010), excess use still is an issue in the public debate. Although the available data limit the opportunity of addressing this question, by comparing the effect of fear of job loss across individuals of different health status, we try to shed some light on it. Our results do not indicate that individuals in good health react much stronger to job worries. Under the assumption of less healthy individuals being in greater need of medical care, this indicates that job insecurity does not discriminate against treatments of limited medical value. Hence, we cannot rule out that treatments of substantial medical benefit are forgone because of fear of unemployment.

In view of this, the prime implication for health policy is to remove the unfavorable signals associated with participation in medical rehabilitation. Germany should modify the existing system of application and acceptance to ensure that the benefits of approved rehabilitation measures justify the associated costs. If rehabilitation measures are actually beneficial and warranted on medical grounds, employers will not misinterpret them as shirking work. Likewise, substantial co-payments for rehabilitations measures should increase the employers' acceptance of the decision of employees to go to rehabilitation. Another option could be to compensate the employer for costs associated with rehabilitation leaves of their employees. Further research may make a valuable contribution by determining the participants' willingness to pay for rehabilitation measures and comparing them to actual costs. This would allow to identify whether and to what extent excess use of medical rehabilitation occurs.

APPENDIX

Table A1. Rehab participation: estimated coefficients (alternative model specifications)

 Reduced modelaBasic model with risk aversionbAlternative health proxy variablesc
 CoefficientS.E.CoefficientS.E.CoefficientS.E.
Structural equation: rehabilitation participation
Fear of unemployment−0.782***0.110−0.826***0.109−0.795***0.102
Age0.0010.0030.0010.0030.006*0.003
Male0.0420.0360.0670.0420.086*0.034
Immigrant0.0450.0750.0970.0760.0490.077
Years of education−0.0140.008−0.0140.009−0.018*0.008
Married0.0050.0450.0260.0500.0040.047
Living with partner0.0120.0500.0040.0590.0130.052
Household size−0.0130.020−0.0240.023−0.0180.020
Number of kids younger than 18 years18 0.0370.0240.0330.0270.048*0.023
Personal income (lag)0.0000.013−0.0030.013−0.0050.013
Occupation
Blue-collar skilled−0.0610.046−0.0080.055−0.0190.047
White-collar low-skilled−0.228***0.045−0.175***0.049−0.214***0.044
White-collar high-skilled−0.348***0.068−0.351***0.075−0.354***0.067
Tenure−0.006**0.002−0.008***0.002−0.006**0.002
Mini-job−0.582***0.068−0.498***0.080−0.576***0.067
Midi-job−0.165*0.074−0.0750.146−0.164*0.073
Temporary work contract0.0650.0850.141*0.055
Side job−0.0610.060−0.172*0.077−0.0930.059
Unemployed (lag)0.0240.0750.0390.0790.0250.073
Firm size
Medium0.178**0.0600.180**0.0690.166**0.061
Large0.221***0.0640.240**0.0800.201**0.065
Year 20040.0340.0430.075*0.0300.0480.041
Year 2006−0.0450.041−0.0210.038
Self-assessed health (lag)
Poor−0.0410.109−0.0620.124
Satisfactory−0.340**0.111−0.289*0.126
Good−0.588***0.120−0.494***0.138
Very good−0.918***0.137−0.848***0.167
Degree of occupational disability0.2590.1720.2790.223
Risk aversion
Medium−0.0200.043
Low−0.180**0.064
Doctor visit (lag)0.180***0.046
Number of doctor visits (lag)0.037***0.009
Hospital visit (lag)0.0580.072
Number of hospital visits (lag)0.100**0.031
Instrumental equation: fear of unemployment 
County unemployment2.697***0.6772.389***0.6612.564***0.660
Age−0.006***0.001−0.005**0.002−0.0020.001
Male0.057*0.0250.076*0.0300.054*0.025
Immigrant0.219***0.0450.202***0.0500.228***0.044
Years of education−0.0100.005−0.0120.006−0.013*0.006
Married0.0430.0350.0480.0400.0470.035
Living with partner0.0130.0360.0310.0420.0390.037
Household size0.0100.014−0.0010.017−0.0030.014
Number of kids younger than 18 years0.0100.0190.0160.0210.0260.019
Personal income (lag)−0.0100.009−0.0100.010−0.0080.009
Occupation
Blue-collar skilled−0.0240.0340.0280.0390.0140.035
White-collar low-skilled−0.210***0.033−0.191***0.035−0.191***0.033
White-collar high-skilled−0.317***0.045−0.298***0.052−0.309***0.046
Tenure−0.007***0.001−0.008***0.002−0.007***0.001
Mini-job−0.496***0.048−0.454***0.054−0.481***0.048
Midi-job−0.202***0.042−0.196*0.097−0.205***0.042
Temporary work contract−0.0890.048−0.090*0.038
Side job−0.084*0.0380.202***0.0430.183***0.037
Unemployed (lag)0.076*0.0370.0400.0470.0480.037
Firm size
Medium0.129***0.0390.120*0.0470.124**0.039
Large0.113**0.0410.111*0.0470.106*0.041
Year 20040.114***0.0200.097***0.0230.112***0.020
Year 20060.0060.0260.0060.026
Self-assessed health (lag)
Poor0.0590.104−0.0010.118
Satisfactory−0.1000.101−0.1190.115
Good−0.238*0.102−0.247*0.118
Very good−0.455***0.104−0.480***0.125
Degree of occupational disability−0.222*0.095−0.229*0.111
Risk aversion
Medium−0.068**0.026
Low−0.140***0.041
Doctor visit (lag)0.055*0.022
Number of doctor visits (lag)0.0050.004
Hospital visit (lag)−0.0860.066
Number of hospital visits (lag)0.0640.046
State dummiesYes***Yes***Yes*** 
ρ0.794***0.1070.846***0.1030.812***0.098
Threshold values ordered probit (instrumental equation)−0.3090.184−0.413*0.2100.0970.164
1.014***0.1840.917***0.2101.413***0.163
Joint significance (x2-statistic)541914844841
Number of observations18 06011 46217 980

*** significant at 1%; ** significant at 5%; * significant at 10%. Standard errors clustered by county.

aThe variable ‘temporary work contract’ is omitted from the basic model specification in order to check whether the variable takes away parts of the effect of fear of unemployment on participation in medical rehabilitation. In the light of the results, this seems not to be the case.

bThe reference category consists of highly risk adverse individuals. The original variable is coded from 0 (no willingness to take risks at all) to 10 (very high willingness to take risks) and is only available for the years 2004 and 2006. Risk attitudes seem to have no impact on the main result. Risk adverse employees are more worried about their job security and have a higher likelihood to participate in medical rehabilitation (both tests of joint significance are significant at the 5% level).

cIncluding alternative health proxy variables yields qualitatively the same results. In quantitative terms, the change in the coefficient of fear of unemployment is marginal. Descriptive statistics for the two risk aversion indicators, ‘doctor visit’, ‘number of doctor visits’, ‘hospital visit’, and ‘number of hospital visits’ are available upon request.

Source: Own calculations.

CONFLICT OF INTEREST

The authors have no conflict of interest.

ACKNOWLEDGEMENTS

The authors are grateful to Ronald Bachmann, Thomas K. Bauer, Colin Green, Katja Görlitz, Maarten Lindeboom, Alfredo R. Paloyo, Christoph M. Schmidt, Hendrik Schmitz, two anonymous referees, and to the participants at the 2010 conferences of the European Society for Population Economics and the Ruhr Graduate School, as well to the participants at the 2011 meeting of the German Health Economics Association for their helpful comments. We also thank Adam Pilny for research assistance. The data used in this paper were extracted using the Add-On package PanelWhiz for StataR. PanelWhiz (http://www.PanelWhiz.eu) was written by John P. Haisken-DeNew (john@PanelWhiz.eu). See Haisken-DeNew and Hahn (2006) for details. The PanelWhiz generated DO file to retrieve the data used here is available from the authors upon request. The paper contains original unpublished work and is not submitted for publication elsewhere. The research is funded by the RWI.

  1. 1

    Pauly et al. (2002) and Nicholson et al. (2006) show that a worker's wage is the lower bound estimate of the firm's cost of sickness absence. On the other hand, on-the-job employee illness causes considerable costs to the firm, too (Pauly et al., 2008).

  2. 2

    Randomized analyses on the effectiveness of medical rehabilitation treatments are still scarce. However, focusing on specific disorders, they unanimously yield significant short-term health improvements, for example, Walsh et al. (1991) and Goossens et al. (1998). To our knowledge, there is no reliable analysis on long-term health effects at all.

  3. 3

    Depending on the case, other authorities are considered as well. The conditions vary from authority to authority.

  4. 4

    For individuals holding private health insurance, coverage for rehabilitation depends on the individual contract.

  5. 5

    Small employers with less than 30 employees are subject to a mandatory ‘employers insurance’, which typically reimburses 80% of the sick pay. Unemployed patients are entitled to a so-called interim payment, which they have to claim against the relevant reimbursement authority.

  6. 6

    Up to an income of €1200 per month, participants face reduced co-payments with increasing exemption depending on their gross income level.

  7. 7

    The share of expenditures for medical rehabilitation treatments amounts to 3.1% of total German health care expenditures (Augurzky et al., 2011). Our own rough estimate of the fraction of sickness absence caused by rehabilitation participation amounts to 8% (there are no specific official data).

  8. 8

    The SOEP addresses rehabilitation participation in the waves of 2004, 2005, and 2007. The relevant questions are not linked to contemporaneous rehabilitation participation. Rather, they are phrased in terms of ‘receiving treatment toward medical rehabilitation, last (calendar) year’.

  9. 9

    Models with alternative variables measuring job insecurity are also estimated, yielding qualitatively similar results. For instance, interviewees are asked about the self-assessed likelihood of losing their jobs within the next 2 years, measured at the 0–100 interval. For a detailed discussion on the concept of job insecurity in the light of the SOEP, see Geishecker (2010).

  10. 10

    Counties exhibit pronounced heterogeneity with respect to both. This is partly due to the pattern that large and medium-sized cities and towns typically constitute one county (urban county, ‘Stadtkreis’), and their surrounding countryside constitutes another (rural county, ‘Landkreis’). Thus, ample rural counties often border directly on densely populated urban counties. The two largest German cities, Berlin and Hamburg, constitute urban counties as well. County populations, hence, range from less than 35,000 to almost 3.5 million people.

  11. 11

    An endogeneity bias may stem from health and income effects of participating in medical rehabilitation. Lagged values cannot be influenced by contemporaneous rehabilitation participation.

  12. 12

    Klosterhuis (1994) finds that rehabilitation utilization varies over regions in Germany.

  13. 13

    The reason is that there is still sufficient variation in job insecurity conditional on temporary work contract to identify the effect of interest.

  14. 14

    The results are also robust with respect to the inclusion of dummy variables for the sector of occupation (NACE at 2-digit level).

  15. 15

    In Germany, public servants are subject to public law and special obligations such as exercising their office on behalf of the common good and serving in a relationship of loyalty. They are permanently employed but prohibited from going on strike. By contrast, public sector employees are employed on the basis of a contract under private law, which applies to all employees in Germany. Their specific working conditions are set out in collective agreements negotiated between the public employers and labor unions (FMI, 2007).

  16. 16

    This applies to roughly 3% of the observations in our sample. All results in this paper are robust to including them into the estimation.

  17. 17

    Surprisingly, there are still elderly people in our sample (Table 2). This is due to a few pensioners holding down a job in order to receive additional income.

  18. 18

    In Section 4, we assess the validity of the instrumental variable, finding support for this line of argument.

  19. 19

    This is warranted by the absence of interaction effects between county unemployment and being a worker that lives more than 5 km away from the workplace (commuting worker) as well as between county unemployment and commuted kilometers in the reduced form. Moreover, our results are robust with respect to the exclusion of commuting workers.

  20. 20

    We use the bioprobit Stata ado-file (Sajaia, 2008). A special feature of this model is that Equation (1) can neither be directly estimated nor jointly with Equation (2) as the regressor math formula is unobserved. Hence, we cannot report estimation results for the model that ignores endogeneity of math formula.

  21. 21

    Running an instrumental variable estimation of the linear probability model with a binary variable for job insecurity (e.g., indicating whether employees are somewhat or very concerned about the security of the job) yields qualitatively the same results. The first stage results are very similar in terms of the sign of the coefficients and significance levels. For instance, the coefficient of the county unemployment rate is strongly positive and significant.

  22. 22

    Clustering at the county level also accounts for the panel structure of the data, as clustered standard errors are robust to arbitrary intra-cluster error correlation. Furthermore, clustering by county accommodates inter-temporal error correlation at the individual level. When clustering the standard errors at the individual level, the coefficient of fear of unemployment remains highly significant.

  23. 23

    It corresponds to a the range between the 36th and the 81st percentile of the distribution of A, see Table 2. Moreover, according to the relevant coefficient estimate, a change in math formula of this size – ceteris paribus – corresponds to an increase in unemployment of 50 percentage points, see Table 4.

  24. 24

    The effect is calculated at the individual level as math formula. We rescale by math formula, because conditional on math formula, that is, asymptotically conditional on εAi = 0, εRi is not standard-normal, but εRi ∼ N(0, 1 − ρ2) holds.

  25. 25

    The null hypothesis of the absolute differential in coefficients being smaller than 0.68 still is rejected.

  26. 26

    As, in the present model, the instrumental equation is estimated by ordered probit rather than ordinary least squares, the test on instrument relevance is based on a Wald-test (χ2(1) − statistic : 15.58) rather than an F-test, as suggested by Staiger and Stock (1997). Nonetheless, as just a single instrument is tested and the number of observations is relatively large, both tests almost coincide (Greene, 2000).

  27. 27

    The bootstrapped 0.95-confidence interval ranges from − 2.52 to − 0.51.

  28. 28

    This roughly represents the East–West differential in Germany's regional unemployment figures.

  29. 29

    Moreover, unhealthy individuals may suffer from lifestyle-related health problems, where medical rehabilitation may not be very effective.

  30. 30

    Numerous analyses point to self-assessed health representing a good proxy for the true health status; for example, McCallum et al. (1994) and Idler and Kasl (1995).

  31. 31

    Heterogeneity in point estimates is by far less distinct if relative rather than absolute changes in the probability of rehabilitation utilization are considered. For the stratified estimation, point estimates do not even exhibit a monotonic pattern. According to joint test statistics, the difference in relative marginal effects is not significant irrespective of the considered model.

Ancillary