- Email: [email protected]
The impact of health status on the duration of unemployment spells and the implications for studies of the impact of unemployment on health status
Journal of Health Economics 20 (2001) 781–796
The impact of health status on the duration of unemployment spells and the implications for studies of the impact of unemployment on health status Jennifer M. Stewart∗ Department of Economics, National University of Ireland, Galway, Ireland Received 10 February 1999; received in revised form 28 February 2001; accepted 3 April 2001
Abstract This paper examines the impact of health status on the duration of unemployment spells and finds that individuals with impaired health will have significantly longer unemployment spells. These longer unemployment spells will result in the stock of unemployed being composed of a larger proportion of individuals with impaired health than the stock of employed. Although, this difference in composition between the stock of unemployed and stock of employed can account for some of the observed difference in mortality rates, it cannot explain all of the difference observed in earlier studies. © 2001 Elsevier Science B.V. All rights reserved. JEL classification: I12 Keywords: Health status; Unemployment; Duration analysis
1. Introduction Health policy has traditionally focused on the role of health care in maintaining health status, but the primacy of health care in health policy has been questioned. Other factors such as position in the social hierarchy or social networks might be equally, or even more, important. Health status is too complicated to be explained by a simple model of health care. One socioeconomic factor that has been examined for its impact on health status has been unemployment. A causal relationship between unemployment and health status has implications for economic policy. Any decision made by the government that has an impact on the unemployment rate requires an understanding of the full social costs of unemployment. If an increase in the unemployment rate reduces health status then the ∗ Tel.: +353-91-52441; fax: +353-91-524130. E-mail address: [email protected] (J.M. Stewart).
0167-6296/01/$ – see front matter © 2001 Elsevier Science B.V. All rights reserved. PII: S 0 1 6 7 - 6 2 9 6 ( 0 1 ) 0 0 0 8 7 - X
782
J.M. Stewart / Journal of Health Economics 20 (2001) 781–796
change in health status needs to be considered when making decisions. Previous papers have examined this issue with both aggregate and microlevel data sets using a variety of econometric techniques. Using a unique data set, I instead examine the effects of health status on unemployment, specifically the impact of impaired health on the duration of unemployment spells. The implication for the results reached in previous work is that selection bias into the stock of unemployed may be an important problem which has led to overestimates of the impact of unemployment on health. This selection bias needs to be dealt with in order to accurately measure the social cost of unemployment. The methodology of this paper proceeds in two stages. First, the impact of impaired health on the duration of unemployment spells is estimated using a duration analysis framework. Second, given that there is a difference in the duration of unemployment spells by health status, the implied difference in the composition of the stock of unemployed and the stock of employed is estimated. The results show that impaired health significantly increases the length of unemployment spells. On this basis alone, therefore, the fraction of the stock of unemployed with impaired health is greater than the fraction of the stock of employed. These results highlight the selection bias into the stock of unemployed that needs to be considered when measuring the impact of unemployment on health status.
2. Background The literature on the relationship between health and unemployment consists of two main streams of research. The seminal paper on unemployment and health by Brenner (1979) motivated the first stream of research which is characterized by the use of aggregate data to estimate the correlation among aggregate variables. Brenner found that current and lagged unemployment rates had a significant impact on mortality rates. Other papers in this stream of research include McAvinchey (1984, 1988), Joyce (1989), Junankar (1991), Joyce and Mocan (1993), Ruhm (1996), Yang and Lester (1995), Hemstrom (1999), Magnusson et al. (1999). Wagstaff (1985) identifies many of the shortcomings of this research. The research undertaken subsequent to Brenner has produced contradictory results. There is some evidence in this literature that, if unemployment does increase mortality rates, there is a long lag until the impact is apparent. The second stream of research has used microlevel data to examine the impact of unemployment on mortality. These studies use a measure of unemployment status at a point in time for their sample which they then use to compare the subsequent mortality rates. These studies find that individuals who were unemployed have a higher subsequent mortality rate than those who were employed, even after controlling for characteristics such as social class, region, age, marital status, sex and occupation. The papers in this stream of research include Moser, Fox and Jones (1984), Moser et al. (1986, 1987), Iversen et al. (1987), Martikainen (1990), Morris et al. (1994), Bartley and Owen (1996), Wadsworth et al. (1999). A potential problem in this second stream of research, which all the authors acknowledge, is that of health-related selection into the stock of unemployed. There are two possible reasons that the stock of unemployed may contain individuals with lower health status than the stock of employed other than the possibility that unemployment decreases health status. The first reason is that individuals with lower health status may be more likely to become
J.M. Stewart / Journal of Health Economics 20 (2001) 781–796
783
unemployed. Arrow (1996) examines this possibility. This paper’s results indicate that for some categories of workers health selection into unemployment occurs. Montgomery et al. (1996) find that for young men the probability of experiencing unemployment is positively related to accumulated health risks throughout childhood. Martikainen and Valkonen (1996, 1998) show that as the unemployment rate rises the relative mortality rate between the unemployed and the employed decreases. They argue that this result occurs because of health selection into unemployment, that is, the unhealthy are likely to become unemployed first and then as the unemployment rate increases the people becoming newly unemployed are healthier. These papers supply evidence that health selection into unemployment does occur. A second reason that the unemployed may be less healthy than the employed is that individuals with lower health status may be more likely to remain unemployed, that is, they may have longer unemployment spells. If individuals in poor health do have longer unemployment spells, then they will comprise a larger proportion of the stock of unemployed than of the stock of employed. It is this second reason that is examined in this paper. Do individuals with impaired health have longer unemployment spells and, if so, what is the impact on the stock of the unemployed?
3. Methodology The analysis of unemployment spells starts by assuming that individuals will leave unemployment if the value of employment is larger than the value of remaining on unemployment. The value of employment is a function of net earnings, that is the difference between earnings and cost of employment, wi , personal characteristics, zi , and the length of time spent on unemployment, T U , or ViE = V E (wi , zi , T U )
(1)
Net earnings increase the value of employment. Personal characteristics could have a direct and indirect impact on the value of employment. The direct impact is that preferences between labor and leisure could be affected by personal characteristics. The indirect impact is that personal characteristics affect an individual’s net earnings. Personal characteristics are predicted to have the same indirect impact on earnings as they do directly on the value of employment. For example, education would have a positive impact on net earnings and it would also make labor more attractive relative to leisure. The amount of time spent unemployed decreases the value of employment because skills may atrophy while individuals are unemployed thereby decreasing their potential earnings. The value of remaining unemployed is a function of personal characteristics, zi , and the length of time spent unemployed, T U , or ViU = V U (zi , T U )
(2)
Personal characteristics have the opposite impact than they do on the value of employment. For example, education would decrease the value of unemployment. The length of time unemployed would increase the value of remaining unemployed because the cost individuals bear from the stigma of being unemployed may decrease as time passes.
784
J.M. Stewart / Journal of Health Economics 20 (2001) 781–796
The hazard rate for unemployment spells is defined as the probability that an unemployment spell ends at time t given that it lasted until time t. The hazard rate is the probability that at time t the value of employment is greater than the value of unemployment times the probability that there is an offer of employment, v λ(wi , zi , T U ) = v Prob[V E (wi , zi , T U ) > V U (zi , T U )]
(3)
If V E or v increase or V U decreases then the hazard rate is predicted to increase. Any variable which affects both V E and V U is predicted to have an opposite impact on each value and therefore an unambiguous impact on the hazard rate. For example, education will increase the value of employment and decrease the value of unemployment. The probability that value of employment is greater than the value of unemployment will increase and the hazard rate will increase. The Prentice–Gloeckler–Meyer (PGM) piecewise constant proportional hazard approach is used to obtain estimates of the baseline hazard function and coefficients. The advantages of this procedure over the Cox proportional hazard model are that relatively few assumptions about the shape of the baseline hazard function are necessary and it is possible to control for unobserved heterogeneity assuming it follows a gamma distribution. Imposition of incorrect assumptions about the baseline and an absence of controls for unobserved heterogeneity could bias the estimated coefficients. However, there are limitations to this approach and other proportional hazard models. One limitation is that the shape of the baseline hazard function is assumed to be the same for everyone. A related limitation is that the covariates are assumed to shift the baseline proportionately but not to affect its shape. The proportional hazard model with no unobserved heterogeneity is specified as hi (t, zi , β) = h0 (t) exp(zi β)
(4)
where hi is individual i’s hazard rate, zi a vector of individual i’s characteristics, β the parameter vector to be estimated, and h0 (t) the baseline hazard function to be estimated. The baseline hazard is divided into intervals. A hazard rate is estimated for each interval and is assumed to be constant over the interval. The decision concerning the number and length of the intervals must balance a desire for functional flexibility (more intervals) with convergence time and estimator precision (fewer intervals). My choice has been guided both by the specifications of previous researchers and by my own empirical hazard functions. The baseline hazard is divided into 51 intervals. For a detailed derivation of the log-likelihood functions, for specifications with and without unobserved heterogeneity, see Prentice and Gloeckler (1978), Meyer (1987, 1990) and Lancaster (1990). The probability that person i’s spell does not end in the j th interval, given that it lasted until the beginning of that interval, is tj Prob[Ti ≥ tj |Ti ≥ tj −1 ] = exp − hi (u, zi , β) du = exp[−exp(zi β + γ (tj ))] tj −1
tj
(5)
where γ (tj ) = ln tj −1 h0 (u) du and Ti is individual i’s spell length. The survival function at time ti , the end of the interval that contains Ti , is the product of the probabilities in Eq. (5) from tj = 1 to ti and is denoted as S(ti ; zi ). The probability of individual i having
J.M. Stewart / Journal of Health Economics 20 (2001) 781–796
785
an uncensored spell that ends in interval i, given that the spell is uncensored, is S(ti−1 ; zi ) − S(ti ; zi )
(6)
Given that a spell is to be censored at interval i, the probability of individual i surviving to interval i is S(ti−1 ; zi )
(7)
Using Eqs. (6) and (7), the log-likelihood function can be written as L(γ , β) =
N i=1
=
N i=1
log(δi (S(ti−1 ; zi ) − S(ti ; zi )) + (1 − δi )S(ti−1 ; zi )) δi log(1 − exp(−exp(γ (ki ) + zi β))) −
ki−1 tj =1
exp(γ (tj ) + zi β) (8)
where δi = 1 if not right censored, ki = min[ti , Ci ], and Ci is the censoring time. The log-likelihood function can be constructed by conditioning on right censoring because it is assumed that censoring is exogenous. This method can be modified to control for one distribution of unobserved heterogeneity. By assuming that the unobserved heterogeneity takes a multiplicative gamma form, the hazard function becomes hi (t, zi , β, θi ) = θi h0 (t) exp(zi β)
(9)
where θi is a random variable that is assumed to be independent of zi and follows a gamma distribution with a mean normalized to one and a variance of σ 2 . A gamma distributed variable has a non-negative support which gives a closed form expression for the likelihood function. The log-likelihood function can then be written as ki−1 N log exp −θ exp(γ (tj ) + zi β) dµ(θ) L(γ , β, µ) = i=1
−δi
tj =1
ki
exp −θ
tj =1
exp(γ (tj ) + zi β) dµ(θ)
(10)
The log-likelihood function is obtained by conditioning on the unobserved θ and integrating over its distribution. The log-likelihood then becomes −1/σ 2 ki−1 N log 1 + σ 2 exp(γ (tj ) + zi β) L(γ , β, σ 2 ) = i=1
tj =1
−δi 1 + σ 2
ki
tj =1
−1/σ 2 exp(γ (tj ) + zi β)
(11)
786
J.M. Stewart / Journal of Health Economics 20 (2001) 781–796
The estimated coefficients from the PGM procedure will give an estimate of the impact of impaired health on the duration of unemployment controlling for other socioeconomic characteristics (see Stewart, 1999). The next step, given that individuals with impaired health status have longer unemployment spells, is to determine the composition of the stock of the unemployed by health status. Begin by considering the flow into unemployment for group i. Group i’s share of the flow into unemployment is denoted by flowi and is equal to flowi =
Ni N
(12)
where N denotes the number of individuals that are entering unemployment and the subscript i denotes the group. Group i’s share of the flow into unemployment is equal to the number of individuals in group i entering unemployment as a proportion of the total number of individuals entering unemployment. Next, define group i’s share of the stock of unemployment as stock i and equal to stock i =
Ui U
(13)
where U refers to the number of unemployed. In the steady state, where the flows into and out of unemployment are constant, it can be shown that U is equal to the sum, over the groups, of the number of individuals entering unemployment from the ith group times the average duration of the ith group’s spells. Combining the definitions for the flow and stock shares, Eqs. (12) and (13), and assuming the steady state gives stock i = flowi
Di D
(14)
where D refers to the average completed duration of the unemployment spells. In the steady state, group i’s share of the stock of unemployed is equal to their share of the flow into unemployment times the ratio of the group i’s average duration to the overall average duration. The average duration can be derived using the estimates from the PGM procedure.
4. The data set The data are from the Canadian Out of Employment Panel (COEP) 1995 survey which was commissioned by Human Resources Development Canada (HRDC). The COEP 1995 survey sample is composed of individuals who received a Record of Employment (ROE) during either one of two periods, 29 January–1 April, cohort 1, or 23 April–3 June, cohort 2, and had a social insurance number ending in “5”. A ROE must be submitted by the employer to HRDC whenever a job termination occurs and allows the beginning of any unemployment spell to be identified, therefore, there are no left censored spells in the data. Individuals were interviewed two times (at approximately 36 and 62 weeks) after the job termination for which the ROE was issued. The two interviews allowed the actual duration of unemployment to be determined up to a maximum of 62 weeks at which point the data was right censored. The data set was merged with unemployment insurance (UI) administrative data and income tax records for the previous 5 years.
J.M. Stewart / Journal of Health Economics 20 (2001) 781–796
787
This paper focuses on the unemployed in the survey, therefore, the sample was restricted to include only individuals who would be included in the stock of unemployed at a point in time in order to be comparable to previous studies. With this sample paring, the remaining sample can be thought of as a sample of the flow into unemployment. Individuals who indicated that they exited the labor force at the time the ROE was issued (therefore, not unemployed) were excluded from the study. Individuals excluded on this basis include the Table 1 Description of variables Variable
Description
Age Not married Less than high school High school Child under 6 Visible minority
Age in 1995 1 if never married, divorced, separated, and widowed, 0 if married or common-law 1 if completed less than high school 1 if at least completed high school, but no university education 1 if there is at least one child under the age of six living in the household 1 if the individual answered yes to the question “By virtue of your ethnic origin are you a visible minority in Canada”? Average of the individual’s total reported income over the previous 5 years. In the hazard models the natural logarithm of this variable is used 1 if the individual was in cohort 2; cohort 1 were jobs that ended between 29 January and 1 April and cohort 2 were jobs that ended between 23 April and 3 June 1 if the ROE job was a not full time position 1 if the individual was covered by a union contract in the ROE job if the individual did not expect to return to the ROE job when it ended if the individual had no UI claims in the previous 5 years if the ROE job was seasonal if occupation is in the professional category which includes, for example, officials and administrators, government, other managers and administrators, architects and engineers, architecture and engineering-related, elementary, secondary and related, other teaching and related, health diagnosing and treating if occupation is in the clerical category which includes, for example, stenographic and typing, bookkeeping, account recording, and related if the occupation is in the services category which includes sales, commodities, sales, services and other sales if the occupation is in the primary industry category which includes, for example, farmers, fishing, hunting, trapping, forestry and logging, mining and quarrying if the occupation is in the industrial category which includes, for example, food, beverage and related, other processing occupations, metal shaping and forming occupations, other machining occupations, metal products,plastics and other related, mechanics and repairmen, motor transport operators if the individual answered yes to the question “Are you limited in the kind or amount of activity that you can do at work because of a long-term physical condition, mental condition or health problem”? if individual left ROE job because of illnes or injury if individual quit ROE job if individual was laid off from ROE job if individual was dismissed or fired from ROE job All results included controls for the region in which the individual was residing; the four regions were the Atlantic provinces, Quebec, Ontario, and the Western provinces
Income Cohort 2 Not full time Union contract No expectation to return No previous UI claim Seasonal job Professional
Clerical Service Primary industry Industrial
Health limitation
Ill Quit Laid off Dismissed Region
788
J.M. Stewart / Journal of Health Economics 20 (2001) 781–796
retired, students, homemakers, and women on maternity leave (824 individuals). Individuals who did not have a spell of unemployment were also excluded (722 individuals) which would include individuals changing jobs. The results presented are conditional on having a spell of unemployment. Individuals were also excluded if they were over the age of 64 (17 individuals) or if their average income over the last 5 years was not between CAD$ 0 and 150,000 (20 individuals). A further 476 individuals were excluded because of missing values of variables used in the duration analysis. Individuals living in the North West Territories or the Yukon were excluded due to small sample size. The final sample included 5817 individuals. Two variables in the data set were used to proxy for health status. First, the main reason the ROE job ended was asked in the interview. One potential reason for the job termination was illness or injury. Having a job end because of illness or injury indicates an impaired health status. Second, individuals were asked if they were “limited in the kind or amount of activity that [they] can do at work because of a long-term physical condition, mental condition or health problem”. A limitation of this kind would also indicate a health impairment. Bartley and Owen (1996) use a similar indicator. Both of these measures of health status directly relate to an individual’s ability to work and so need to be included in an analysis of unemployment duration. These measures will, however, miss individuals with a lower health status that does not affect their ability to work (Table 1). Table 2 Socioeconomic characteristics of total sample and by reason for leaving ROE job (enteries are in percentages unless otherwise indicated) Variable (years)∗
Age Not married∗ Less than high school∗ High school Female∗ Child under 6 Visible minority∗ Income × 1000 (CAD$)∗ Cohort 2∗ Not full time∗ Union contract∗ No expectation to return∗ No previous UI claim∗ Seasonal job∗ Professional∗ Clerical∗ Service∗ Primary industry∗ Industrial∗ Limited∗ Spell length∗ (Weeks) Sample size
Total
Illness
Quit
Laid off
Dismissed
36 39 26 31 39 19 19 22 52 28 32 46 41 28 9 3 6 5 77 8 24 5817
40 34 34 32 55 19 23 21 51 30 31 21 43 9 9 2 4 13 72 37 34 334
31 48 17 30 41 20 16 20 56 34 14 88 77 11 12 4 2 4 78 5 18 1108
28 36 28 31 38 19 20 23 51 27 37 35 31 35 8 3 7 4 78 6 25 4198
33 58 28 28 34 18 17 18 40 25 11 84 56 7 10 2 0 5 83 11 33 177
∗ Indicates that a joint t-test of the equality of the means across the reasons was rejected at a 5% significance level.
J.M. Stewart / Journal of Health Economics 20 (2001) 781–796
789
Table 2 presents the composition of the sample by the reason the ROE job ended and by the variables used in the empirical analysis. An asterisk beside a variable name indicates that the joint t-test that the means for each reason for termination are equal was rejected at the 5% significance level. The only variables for which the joint test was not rejected are the proportion of the group that have completed high school and the proportion of the group that have a child less than 6 years old. The reported spell length accounts for censoring by estimating an expected duration for censored spells and assigning the estimated spell length to any censored spells. Individuals who report leaving the ROE job because of illness appear to be older, less educated, and more likely to be female than other categories of individuals. In order to accurately measure the impact of limited health status on the duration of unemployment, it is necessary to extend the analysis of the spell lengths beyond the empirical hazard rate and to control for the impact of these socioeconomic variables. Table 3 compares the socioeconomic composition of individuals who report they have a health limitation and those who report they do not. An asterisk beside a variable name indicates that the difference between the two groups is significantly different from zero at
Table 3 Socioeconomic characteristics of total sample and by presence of a health limitation (enteries are in percentages unless otherwise indicated) Variable
Total
Health limitation
No health limitation
Total sample (%) Age (years)∗ Not married Less than high school∗ High school Female Child under 6 Visible minority Income × 1000 (CAD$)∗ Cohort 2 Full time Union contract No expectation to return No previous UI claim Seasonal job Professional Clerical∗ Service Primary industry Industrial Reason: Ill∗ Quit∗ Laid off∗ Dismissed Spell length∗ (weeks) Sample size
100 36 39 26 31 39 19 19 22 52 72 32 46 41 28 9 3 6 5 77 6 19 72 3 24 5817
8 40 38 36 31 42 19 23 20 47 70 29 38 44 25 7 1 7 6 79 27 12 57 4 35 457
92 36 39 25 31 39 19 19 22 52 72 32 47 41 28 9 3 6 5 77 4 20 73 3 23 5360
∗ Indicates that a joint t-test of the equality of the means across the reasons was rejected at a 5% significance level.
790
J.M. Stewart / Journal of Health Economics 20 (2001) 781–796
a 5% significance level using a t-test. Individuals who report a limitation are older, less educated, have a lower income, are more likely to have left the ROE job because of illness and less likely to have quit or been laid off than those who do not report a limitation. About 27% of individuals with a limitation left their ROE job because of illness or injury compared to only 4% of those without a limitation. 5. Results 5.1. Hazard estimation Table 4 presents the results from four specifications of the hazard function. Two functional forms of the independent variables for health status are used. The first functional form includes interaction terms of the two health status indicators. The interaction terms are included because there are a number of individuals who reported both leaving the ROE job because of illness and having a health limitation. The second functional form includes separate indicators for the presence of a health limitation and for the reason for termination. The omitted category in both functional forms is an individual who does not have a health limitation and who was laid off. Each functional form is used in specifications that do and do not control for unobserved heterogeneity. The coefficients in Table 4 are the exponentials of the actual coefficients estimated and represent the proportional shift in the baseline hazard due to a change in the variable. A value greater than one indicates an upward shift, or higher hazard rate, while a value less than one indicates a downward shift, or lower hazard rate. Having a higher hazard rate indicates that the individual is more likely to leave unemployment and, therefore, expected to have a shorter unemployment spell and vice versa for a lower hazard rate. The estimated coefficients for the explanatory variables are similar across the two forms of incorporating the health status variables, but they do differ, in some cases, depending on whether or not the specification controlled for unobserved heterogeneity. In all four models, the baseline hazard rate refers to the an individual with the same characteristics and so it is not surprising that including the health status variables as interaction terms rather than separate dummy variables does not affect the estimated coefficients for other explanatory variables. Unobserved heterogeneity can be thought of as an omitted variable, therefore, when it is controlled for it would be expected that coefficients would change if they are correlated with the missing information. The estimated coefficients are inline with expectations. For example, older individuals are predicted to have lower hazard rates, or longer unemployment spells. However, when controlling for unobserved heterogeneity, it is only the two older age groups that have statistically significant lower hazard rates, rather than a consistent difference across all age groups. The specifications that control for unobserved heterogeneity indicate that it exists and, therefore, the preferred specifications are those which control for unobserved heterogeneity. The coefficients of the variables for health status are fairly consistent across specifications. In the specifications which use the interaction terms, having a health limitation is associated with a lower hazard rate compared to those without a health limitation and left the ROE job
J.M. Stewart / Journal of Health Economics 20 (2001) 781–796
791
792
J.M. Stewart / Journal of Health Economics 20 (2001) 781–796
J.M. Stewart / Journal of Health Economics 20 (2001) 781–796
793
for the same reason. Many of the coefficients for the interaction terms are not statistically significant, but individuals with a health limitation and left the ROE job because of illness have a statistically significant lower hazard rate and those who quit the ROE job have a statistically significant higher hazard rate. The results are similar when separate dummy variables are used for the health status indicators. Individuals with a health limitation have a statistically significant lower hazard rate. Leaving the ROE job because of illness is also associated with a statistically significant lower hazard rate while quitting the ROE job is associated with a statistically significant higher hazard rate. These results indicate that impaired health status leads to a lower probability of leaving unemployment and, therefore, longer unemployment spells. 5.2. Effect of hazards on composition of stock of the unemployed Some simplifying assumptions must be made in order to examine the impact of longer unemployment spells of individuals with impaired health status on their share of the stock of unemployed. First, assume a steady state, that is, the composition of the flow into unemployment is constant over time. Second, assume that all individuals are homogenous except for their health status which is classified into eight groups; each of the four reasons for a job termination with and without a health limitation. Table 5 shows the proportion of the flow into unemployment, average duration, and implied proportion of the stock of unemployment for each of the eight groups. The top three rows of Table 5 are the groups who are healthy and the bottom five rows have impaired health status. The share of the flow into unemployment for the top three groups combined is 88.54% and for the groups with impaired health status it is 11.46%. When the differences in average duration are taken into account, the shares in the stock of unemployment are 78.55% for the healthy groups and 21.45% for the groups with impaired health status. Although, individuals with lower health status comprised only 11.46% of the flow into unemployment, they account for 21.45% of the stock of unemployed, which is almost double their entrance share, because of their longer unemployment spells. What are the implications for a study that has compared mortality rates of the stock of employed and unemployed? Assume that individuals with high health status have lower mortality rates than those with impaired health status, but that there is no difference in Table 5 Flow composition, average duration, and implied stock composition Group
Flowi
Average duration
Stock i
Laid off no health limitation Quit no health limitation Dismissed no health limitation Illness no health limitation Illness health limitation Dismissed health limitation Quit health limitation Laid off health limitation
67.70 18.14 2.70 3.61 2.13 0.34 0.91 4.47
3.13 0.25 3.53 4.43 12.83 4.54 0.86 3.61
73.68 1.57 3.30 5.56 9.48 0.54 0.27 5.60
794
J.M. Stewart / Journal of Health Economics 20 (2001) 781–796
mortality rates between employed and unemployed individuals. Also, assume that there is no selection bias into the flow into unemployment by health status so that the composition of the stock of employed by health status is the same as the flow out of it. These restrictions imply that the estimate of the impact presented here is a lower bound in some sense. The predicted mortality rate for the stock of employed (MRE ) will be a weighted average of the mortality rate for healthy individuals (MRH ) and the mortality rate for individuals with impaired health (MRIH ) MRE = (0.8854)MRH + (0.1146)MRIH
(15)
where the proportion of the stock in each health status is from the data. The predicted mortality rate for the stock of unemployed (MRU ) is MRU = (0.7855)MRH + (0.2145)MRIH
(16)
where the proportion of the stock in each health status is from the estimated stock share, when the differences in the durations is taken into account. The difference between the mortality rates is MRU −MRE = (−0.0999)MRH + (0.0999)MRIH = (0.0999)(MRIH −MRH )
(17)
Even if unemployment has no impact on health status, the relative mortality rate between the unemployed and the employed will still be greater than one. Can the results in this paper fully explain the difference in mortality rates between the unemployed and employed observed in previous papers? If we examine the lower bound of the estimate from Iversen et al. (1987), which is a relative mortality rate of 1.33, then MRU (0.7855)MRH + (0.2145)MRIH = = 1.33 MRE (0.8854)MRH + (0.1146)MRIH
(18)
which would imply that MRIH = 6.31 MRH
(19)
It would be necessary for the relative mortality rate between individuals with impaired health and those without impaired health to be around 6.3 for the results from this paper to fully explain the lower bound of the previously observed differences. This condition seems unlikely to be met. Those with a health limitation only compose a small part of the flow into unemployment, most individuals enter unemployment because they have been laid off. Even though the longer unemployment spells doubles the proportion of the stock of the unemployed that have a health limitation compared to their proportion of the flow into unemployment, they still only comprise around 20% of the stock of unemployed. In order for this small proportion to have a significant impact on the overall mortality rate of the stock of unemployed, their relative mortality rate to those without a health limitation needs to be high. The simplifying assumptions that have been made must also be kept in mind when considering this result. It has been assumed that the only differences between individuals are the reason they left the ROE job and the presence of a health limitation.
J.M. Stewart / Journal of Health Economics 20 (2001) 781–796
795
The impact of other socioeconomic characteristics which are correlated with limited health status has not been included. Although, health-related selection due to longer unemployment spells does appear to occur and implies that previous research overestimates the impact of unemployment on mortality, it explains only some of the observed difference in mortality rates between the unemployed and employed.
6. Conclusion Numerous studies have found an empirical link between unemployment and mortality rates. One hypothesis is that unemployment decreases health status, but this relationship could be observed for three reasons. First, unemployment does decrease health status. Second, individuals with impaired health status are more likely to become unemployed and, therefore, comprise a larger proportion of the stock of unemployed. Third, individuals with impaired health are more likely to remain unemployed and, therefore, will comprise a larger proportion of the stock of unemployed. This paper focused on the third issue. Are individuals with impaired health more likely to remain unemployed? Duration analysis was used to estimate the effect of health status and other socioeconomic factors on the length of unemployment spells using a longitudinal data set that interviewed individuals after a job separation. Two proxies for impaired health were examined in this paper: a job termination because of illness and a reported health limitation. The PGM procedure was used to estimate a hazard function which allowed for controls for unobserved heterogeneity. A job termination due to illness and a reported health limitation both had a significant, negative impact on the hazard rate. The probability of finding employment at any point in time is lower if the individual quit due to illness or had a health limitation. In particular, individuals who reported both a job termination due to illness and a health limitation had long unemployment spells. The stock of unemployed will be comprised of a larger proportion of individuals with impaired health than the stock of employed because of the longer unemployment spells of these individuals. The implication for previous research is that the estimated impact of unemployment on health status will be an overestimate. To accurately measure the social cost of unemployment the results in this paper, and the results of other researchers, need to be taken into account. References Arrow, J.O., 1996. Estimating the influence of health as a risk factor of unemployment: a survival analysis of employment durations for workers surveyed in the German socioeconomic panel. Social Science and Medicine 42 (12), 1651–1659. Bartley, M., Owen, C., 1996. Relation between socioeconomic status, employment, and health during economic change, 1973–1993. British Medical Journal 313, 445–449. Brenner, M.H., 1979. Mortality and the national economy: a review and the experience of England and Wales, 1936–76. Lancet, September 15, 568–573. Hemstrom, O., 1999. Explaining differential rates of mortality decline for Swedish men and women: a time series analysis, 1945–1992. Social Science and Medicine 48, 1759–1777.
796
J.M. Stewart / Journal of Health Economics 20 (2001) 781–796
Iversen, L., Andersen, O., Andersen, P.K., Christoffersen, K., Keiding, N., 1987. Unemployment and mortality in Denmark, 1970–80. British Medical Journal 295, 878–884. Joyce, T.J., 1989. A time series analysis of unemployment and health: the case of Birth outcomes in New York city. Journal of Health Economics 8 (4), 419–436. Joyce, T.J., Mocan, N., 1993. Unemployment and infant health: time series evidence from the state of Tennessee. Journal of Human Resources 28 (1), 185–203. Junankar, P.N., 1991. Unemployment and mortality in England and Wales: a preliminary analysis. Oxford Economic Papers 43 (2), 305–320. Lancaster, T., 1990. The Econometric Analysis of Transition Data. Cambridge University Press, Cambridge. Magnusson, S., Sigurdsson, J.A., Svardsudd, K., 1999. Indicators of health and well-being in Iceland and Sweden: a comparative study of various indicators concerning standards of living and mortality. Scandinavian Journal of Primary Health Care 17, 139–144. Martikainen, P.T., 1990. Unemployment and mortality among Finnish men, 1981–5. British Medical Journal 301, 407–411. Martikainen, P.T., Valkonen, T., 1996. Excess mortality of unemployed men and women during a period of rapidly increasing unemployment. Lancet 348, 909–912 . Martikainen, P.T., Valkonen, T., 1998. The effects of differential unemployment rate increases of occupation groups on changes in mortality. American Journal of Public Health 88 (12), 1859–1861. McAvinchey, I.D., 1984. Economic factors and mortality: some aspects of the Scottish case, 1950–1978. Scottish Journal of Political Economy 31 (1), 1–27. McAvinchey, I.D., 1988. A comparison of unemployment, income and mortality interaction for five European Countries. Applied Economics 20 (4), 453–471. Meyer, B.D., 1987. Hazard and Markov Chain Models with Applications to Labour Economics. Thesis MIT. Meyer, B.D., 1990. On unemployment insurance and unemployment spells. Econometrica 58 (4), 757–782. Montgomery, S.M., Bartley, M.J., Cook, D.G., Wadsworth, M.E.J., 1996. Health and social precursors of unemployment in young men in Great Britain. Journal of Epidemiology and Community Health 50, 415– 422. Morris, J.K., Cook, D.G., Shaper, A.G., 1994. Loss of employment and mortality. British Medical Journal 308, 1135–1139. Moser, K.A., Fox, A.J., Jones, D.R., 1984. Unemployment and mortality in the OPCS longitudinal study. Lancet, 928–942. Moser, K.A., Fox, A.J., Jones, D.R., Goldblatt, P.O., 1986. Unemployment and mortality: further evidence from the OPCS longitudinal study, 1971–1981. Lancet, 365–366. Moser, K.A., Goldblatt, P.O., Fox, A.J., Jones, D.R., 1987. Unemployment and mortality: comparison of the 1971 and 1981 longitudinal census samples. British Medical Journal 294, 86–90. Prentice, R., Gloeckler, L., 1978. Regression analysis of grouped survival data with an application to breast cancer. Biometrics 34, 57–67. Ruhm, C.J., 1996. Are Recessions Good For Your Health? NBER Working Paper 5570. Stewart, J.M., 1999. Three Essays on Unemployment and Social Assistance. Thesis McMaster University. Wadsworth, M.E.J., Montgomery, S.M., Bartley, M.J., 1999. The persisting effect of unemployment on health and social well-being in men early in working life. Social Science and Medicine 48, 1491–1499. Wagstaff, A., 1985. Time series analysis of the relationship between unemployment and mortality: a survey of econometric techniques and replications of Brenner’s study. Social Science and Medicine 21 (9), 985–996. Yang, B., Lester, D., 1995. Suicide, homicide and unemployment. Applied Economic Letters 2, 278–279.
The impact of health status on the duration of unemployment spells and the implications for studies of the impact of unemployment on health status Jennifer M. Stewart∗ Department of Economics, National University of Ireland, Galway, Ireland Received 10 February 1999; received in revised form 28 February 2001; accepted 3 April 2001
Abstract This paper examines the impact of health status on the duration of unemployment spells and finds that individuals with impaired health will have significantly longer unemployment spells. These longer unemployment spells will result in the stock of unemployed being composed of a larger proportion of individuals with impaired health than the stock of employed. Although, this difference in composition between the stock of unemployed and stock of employed can account for some of the observed difference in mortality rates, it cannot explain all of the difference observed in earlier studies. © 2001 Elsevier Science B.V. All rights reserved. JEL classification: I12 Keywords: Health status; Unemployment; Duration analysis
1. Introduction Health policy has traditionally focused on the role of health care in maintaining health status, but the primacy of health care in health policy has been questioned. Other factors such as position in the social hierarchy or social networks might be equally, or even more, important. Health status is too complicated to be explained by a simple model of health care. One socioeconomic factor that has been examined for its impact on health status has been unemployment. A causal relationship between unemployment and health status has implications for economic policy. Any decision made by the government that has an impact on the unemployment rate requires an understanding of the full social costs of unemployment. If an increase in the unemployment rate reduces health status then the ∗ Tel.: +353-91-52441; fax: +353-91-524130. E-mail address: [email protected] (J.M. Stewart).
0167-6296/01/$ – see front matter © 2001 Elsevier Science B.V. All rights reserved. PII: S 0 1 6 7 - 6 2 9 6 ( 0 1 ) 0 0 0 8 7 - X
782
J.M. Stewart / Journal of Health Economics 20 (2001) 781–796
change in health status needs to be considered when making decisions. Previous papers have examined this issue with both aggregate and microlevel data sets using a variety of econometric techniques. Using a unique data set, I instead examine the effects of health status on unemployment, specifically the impact of impaired health on the duration of unemployment spells. The implication for the results reached in previous work is that selection bias into the stock of unemployed may be an important problem which has led to overestimates of the impact of unemployment on health. This selection bias needs to be dealt with in order to accurately measure the social cost of unemployment. The methodology of this paper proceeds in two stages. First, the impact of impaired health on the duration of unemployment spells is estimated using a duration analysis framework. Second, given that there is a difference in the duration of unemployment spells by health status, the implied difference in the composition of the stock of unemployed and the stock of employed is estimated. The results show that impaired health significantly increases the length of unemployment spells. On this basis alone, therefore, the fraction of the stock of unemployed with impaired health is greater than the fraction of the stock of employed. These results highlight the selection bias into the stock of unemployed that needs to be considered when measuring the impact of unemployment on health status.
2. Background The literature on the relationship between health and unemployment consists of two main streams of research. The seminal paper on unemployment and health by Brenner (1979) motivated the first stream of research which is characterized by the use of aggregate data to estimate the correlation among aggregate variables. Brenner found that current and lagged unemployment rates had a significant impact on mortality rates. Other papers in this stream of research include McAvinchey (1984, 1988), Joyce (1989), Junankar (1991), Joyce and Mocan (1993), Ruhm (1996), Yang and Lester (1995), Hemstrom (1999), Magnusson et al. (1999). Wagstaff (1985) identifies many of the shortcomings of this research. The research undertaken subsequent to Brenner has produced contradictory results. There is some evidence in this literature that, if unemployment does increase mortality rates, there is a long lag until the impact is apparent. The second stream of research has used microlevel data to examine the impact of unemployment on mortality. These studies use a measure of unemployment status at a point in time for their sample which they then use to compare the subsequent mortality rates. These studies find that individuals who were unemployed have a higher subsequent mortality rate than those who were employed, even after controlling for characteristics such as social class, region, age, marital status, sex and occupation. The papers in this stream of research include Moser, Fox and Jones (1984), Moser et al. (1986, 1987), Iversen et al. (1987), Martikainen (1990), Morris et al. (1994), Bartley and Owen (1996), Wadsworth et al. (1999). A potential problem in this second stream of research, which all the authors acknowledge, is that of health-related selection into the stock of unemployed. There are two possible reasons that the stock of unemployed may contain individuals with lower health status than the stock of employed other than the possibility that unemployment decreases health status. The first reason is that individuals with lower health status may be more likely to become
J.M. Stewart / Journal of Health Economics 20 (2001) 781–796
783
unemployed. Arrow (1996) examines this possibility. This paper’s results indicate that for some categories of workers health selection into unemployment occurs. Montgomery et al. (1996) find that for young men the probability of experiencing unemployment is positively related to accumulated health risks throughout childhood. Martikainen and Valkonen (1996, 1998) show that as the unemployment rate rises the relative mortality rate between the unemployed and the employed decreases. They argue that this result occurs because of health selection into unemployment, that is, the unhealthy are likely to become unemployed first and then as the unemployment rate increases the people becoming newly unemployed are healthier. These papers supply evidence that health selection into unemployment does occur. A second reason that the unemployed may be less healthy than the employed is that individuals with lower health status may be more likely to remain unemployed, that is, they may have longer unemployment spells. If individuals in poor health do have longer unemployment spells, then they will comprise a larger proportion of the stock of unemployed than of the stock of employed. It is this second reason that is examined in this paper. Do individuals with impaired health have longer unemployment spells and, if so, what is the impact on the stock of the unemployed?
3. Methodology The analysis of unemployment spells starts by assuming that individuals will leave unemployment if the value of employment is larger than the value of remaining on unemployment. The value of employment is a function of net earnings, that is the difference between earnings and cost of employment, wi , personal characteristics, zi , and the length of time spent on unemployment, T U , or ViE = V E (wi , zi , T U )
(1)
Net earnings increase the value of employment. Personal characteristics could have a direct and indirect impact on the value of employment. The direct impact is that preferences between labor and leisure could be affected by personal characteristics. The indirect impact is that personal characteristics affect an individual’s net earnings. Personal characteristics are predicted to have the same indirect impact on earnings as they do directly on the value of employment. For example, education would have a positive impact on net earnings and it would also make labor more attractive relative to leisure. The amount of time spent unemployed decreases the value of employment because skills may atrophy while individuals are unemployed thereby decreasing their potential earnings. The value of remaining unemployed is a function of personal characteristics, zi , and the length of time spent unemployed, T U , or ViU = V U (zi , T U )
(2)
Personal characteristics have the opposite impact than they do on the value of employment. For example, education would decrease the value of unemployment. The length of time unemployed would increase the value of remaining unemployed because the cost individuals bear from the stigma of being unemployed may decrease as time passes.
784
J.M. Stewart / Journal of Health Economics 20 (2001) 781–796
The hazard rate for unemployment spells is defined as the probability that an unemployment spell ends at time t given that it lasted until time t. The hazard rate is the probability that at time t the value of employment is greater than the value of unemployment times the probability that there is an offer of employment, v λ(wi , zi , T U ) = v Prob[V E (wi , zi , T U ) > V U (zi , T U )]
(3)
If V E or v increase or V U decreases then the hazard rate is predicted to increase. Any variable which affects both V E and V U is predicted to have an opposite impact on each value and therefore an unambiguous impact on the hazard rate. For example, education will increase the value of employment and decrease the value of unemployment. The probability that value of employment is greater than the value of unemployment will increase and the hazard rate will increase. The Prentice–Gloeckler–Meyer (PGM) piecewise constant proportional hazard approach is used to obtain estimates of the baseline hazard function and coefficients. The advantages of this procedure over the Cox proportional hazard model are that relatively few assumptions about the shape of the baseline hazard function are necessary and it is possible to control for unobserved heterogeneity assuming it follows a gamma distribution. Imposition of incorrect assumptions about the baseline and an absence of controls for unobserved heterogeneity could bias the estimated coefficients. However, there are limitations to this approach and other proportional hazard models. One limitation is that the shape of the baseline hazard function is assumed to be the same for everyone. A related limitation is that the covariates are assumed to shift the baseline proportionately but not to affect its shape. The proportional hazard model with no unobserved heterogeneity is specified as hi (t, zi , β) = h0 (t) exp(zi β)
(4)
where hi is individual i’s hazard rate, zi a vector of individual i’s characteristics, β the parameter vector to be estimated, and h0 (t) the baseline hazard function to be estimated. The baseline hazard is divided into intervals. A hazard rate is estimated for each interval and is assumed to be constant over the interval. The decision concerning the number and length of the intervals must balance a desire for functional flexibility (more intervals) with convergence time and estimator precision (fewer intervals). My choice has been guided both by the specifications of previous researchers and by my own empirical hazard functions. The baseline hazard is divided into 51 intervals. For a detailed derivation of the log-likelihood functions, for specifications with and without unobserved heterogeneity, see Prentice and Gloeckler (1978), Meyer (1987, 1990) and Lancaster (1990). The probability that person i’s spell does not end in the j th interval, given that it lasted until the beginning of that interval, is tj Prob[Ti ≥ tj |Ti ≥ tj −1 ] = exp − hi (u, zi , β) du = exp[−exp(zi β + γ (tj ))] tj −1
tj
(5)
where γ (tj ) = ln tj −1 h0 (u) du and Ti is individual i’s spell length. The survival function at time ti , the end of the interval that contains Ti , is the product of the probabilities in Eq. (5) from tj = 1 to ti and is denoted as S(ti ; zi ). The probability of individual i having
J.M. Stewart / Journal of Health Economics 20 (2001) 781–796
785
an uncensored spell that ends in interval i, given that the spell is uncensored, is S(ti−1 ; zi ) − S(ti ; zi )
(6)
Given that a spell is to be censored at interval i, the probability of individual i surviving to interval i is S(ti−1 ; zi )
(7)
Using Eqs. (6) and (7), the log-likelihood function can be written as L(γ , β) =
N i=1
=
N i=1
log(δi (S(ti−1 ; zi ) − S(ti ; zi )) + (1 − δi )S(ti−1 ; zi )) δi log(1 − exp(−exp(γ (ki ) + zi β))) −
ki−1 tj =1
exp(γ (tj ) + zi β) (8)
where δi = 1 if not right censored, ki = min[ti , Ci ], and Ci is the censoring time. The log-likelihood function can be constructed by conditioning on right censoring because it is assumed that censoring is exogenous. This method can be modified to control for one distribution of unobserved heterogeneity. By assuming that the unobserved heterogeneity takes a multiplicative gamma form, the hazard function becomes hi (t, zi , β, θi ) = θi h0 (t) exp(zi β)
(9)
where θi is a random variable that is assumed to be independent of zi and follows a gamma distribution with a mean normalized to one and a variance of σ 2 . A gamma distributed variable has a non-negative support which gives a closed form expression for the likelihood function. The log-likelihood function can then be written as ki−1 N log exp −θ exp(γ (tj ) + zi β) dµ(θ) L(γ , β, µ) = i=1
−δi
tj =1
ki
exp −θ
tj =1
exp(γ (tj ) + zi β) dµ(θ)
(10)
The log-likelihood function is obtained by conditioning on the unobserved θ and integrating over its distribution. The log-likelihood then becomes −1/σ 2 ki−1 N log 1 + σ 2 exp(γ (tj ) + zi β) L(γ , β, σ 2 ) = i=1
tj =1
−δi 1 + σ 2
ki
tj =1
−1/σ 2 exp(γ (tj ) + zi β)
(11)
786
J.M. Stewart / Journal of Health Economics 20 (2001) 781–796
The estimated coefficients from the PGM procedure will give an estimate of the impact of impaired health on the duration of unemployment controlling for other socioeconomic characteristics (see Stewart, 1999). The next step, given that individuals with impaired health status have longer unemployment spells, is to determine the composition of the stock of the unemployed by health status. Begin by considering the flow into unemployment for group i. Group i’s share of the flow into unemployment is denoted by flowi and is equal to flowi =
Ni N
(12)
where N denotes the number of individuals that are entering unemployment and the subscript i denotes the group. Group i’s share of the flow into unemployment is equal to the number of individuals in group i entering unemployment as a proportion of the total number of individuals entering unemployment. Next, define group i’s share of the stock of unemployment as stock i and equal to stock i =
Ui U
(13)
where U refers to the number of unemployed. In the steady state, where the flows into and out of unemployment are constant, it can be shown that U is equal to the sum, over the groups, of the number of individuals entering unemployment from the ith group times the average duration of the ith group’s spells. Combining the definitions for the flow and stock shares, Eqs. (12) and (13), and assuming the steady state gives stock i = flowi
Di D
(14)
where D refers to the average completed duration of the unemployment spells. In the steady state, group i’s share of the stock of unemployed is equal to their share of the flow into unemployment times the ratio of the group i’s average duration to the overall average duration. The average duration can be derived using the estimates from the PGM procedure.
4. The data set The data are from the Canadian Out of Employment Panel (COEP) 1995 survey which was commissioned by Human Resources Development Canada (HRDC). The COEP 1995 survey sample is composed of individuals who received a Record of Employment (ROE) during either one of two periods, 29 January–1 April, cohort 1, or 23 April–3 June, cohort 2, and had a social insurance number ending in “5”. A ROE must be submitted by the employer to HRDC whenever a job termination occurs and allows the beginning of any unemployment spell to be identified, therefore, there are no left censored spells in the data. Individuals were interviewed two times (at approximately 36 and 62 weeks) after the job termination for which the ROE was issued. The two interviews allowed the actual duration of unemployment to be determined up to a maximum of 62 weeks at which point the data was right censored. The data set was merged with unemployment insurance (UI) administrative data and income tax records for the previous 5 years.
J.M. Stewart / Journal of Health Economics 20 (2001) 781–796
787
This paper focuses on the unemployed in the survey, therefore, the sample was restricted to include only individuals who would be included in the stock of unemployed at a point in time in order to be comparable to previous studies. With this sample paring, the remaining sample can be thought of as a sample of the flow into unemployment. Individuals who indicated that they exited the labor force at the time the ROE was issued (therefore, not unemployed) were excluded from the study. Individuals excluded on this basis include the Table 1 Description of variables Variable
Description
Age Not married Less than high school High school Child under 6 Visible minority
Age in 1995 1 if never married, divorced, separated, and widowed, 0 if married or common-law 1 if completed less than high school 1 if at least completed high school, but no university education 1 if there is at least one child under the age of six living in the household 1 if the individual answered yes to the question “By virtue of your ethnic origin are you a visible minority in Canada”? Average of the individual’s total reported income over the previous 5 years. In the hazard models the natural logarithm of this variable is used 1 if the individual was in cohort 2; cohort 1 were jobs that ended between 29 January and 1 April and cohort 2 were jobs that ended between 23 April and 3 June 1 if the ROE job was a not full time position 1 if the individual was covered by a union contract in the ROE job if the individual did not expect to return to the ROE job when it ended if the individual had no UI claims in the previous 5 years if the ROE job was seasonal if occupation is in the professional category which includes, for example, officials and administrators, government, other managers and administrators, architects and engineers, architecture and engineering-related, elementary, secondary and related, other teaching and related, health diagnosing and treating if occupation is in the clerical category which includes, for example, stenographic and typing, bookkeeping, account recording, and related if the occupation is in the services category which includes sales, commodities, sales, services and other sales if the occupation is in the primary industry category which includes, for example, farmers, fishing, hunting, trapping, forestry and logging, mining and quarrying if the occupation is in the industrial category which includes, for example, food, beverage and related, other processing occupations, metal shaping and forming occupations, other machining occupations, metal products,plastics and other related, mechanics and repairmen, motor transport operators if the individual answered yes to the question “Are you limited in the kind or amount of activity that you can do at work because of a long-term physical condition, mental condition or health problem”? if individual left ROE job because of illnes or injury if individual quit ROE job if individual was laid off from ROE job if individual was dismissed or fired from ROE job All results included controls for the region in which the individual was residing; the four regions were the Atlantic provinces, Quebec, Ontario, and the Western provinces
Income Cohort 2 Not full time Union contract No expectation to return No previous UI claim Seasonal job Professional
Clerical Service Primary industry Industrial
Health limitation
Ill Quit Laid off Dismissed Region
788
J.M. Stewart / Journal of Health Economics 20 (2001) 781–796
retired, students, homemakers, and women on maternity leave (824 individuals). Individuals who did not have a spell of unemployment were also excluded (722 individuals) which would include individuals changing jobs. The results presented are conditional on having a spell of unemployment. Individuals were also excluded if they were over the age of 64 (17 individuals) or if their average income over the last 5 years was not between CAD$ 0 and 150,000 (20 individuals). A further 476 individuals were excluded because of missing values of variables used in the duration analysis. Individuals living in the North West Territories or the Yukon were excluded due to small sample size. The final sample included 5817 individuals. Two variables in the data set were used to proxy for health status. First, the main reason the ROE job ended was asked in the interview. One potential reason for the job termination was illness or injury. Having a job end because of illness or injury indicates an impaired health status. Second, individuals were asked if they were “limited in the kind or amount of activity that [they] can do at work because of a long-term physical condition, mental condition or health problem”. A limitation of this kind would also indicate a health impairment. Bartley and Owen (1996) use a similar indicator. Both of these measures of health status directly relate to an individual’s ability to work and so need to be included in an analysis of unemployment duration. These measures will, however, miss individuals with a lower health status that does not affect their ability to work (Table 1). Table 2 Socioeconomic characteristics of total sample and by reason for leaving ROE job (enteries are in percentages unless otherwise indicated) Variable (years)∗
Age Not married∗ Less than high school∗ High school Female∗ Child under 6 Visible minority∗ Income × 1000 (CAD$)∗ Cohort 2∗ Not full time∗ Union contract∗ No expectation to return∗ No previous UI claim∗ Seasonal job∗ Professional∗ Clerical∗ Service∗ Primary industry∗ Industrial∗ Limited∗ Spell length∗ (Weeks) Sample size
Total
Illness
Quit
Laid off
Dismissed
36 39 26 31 39 19 19 22 52 28 32 46 41 28 9 3 6 5 77 8 24 5817
40 34 34 32 55 19 23 21 51 30 31 21 43 9 9 2 4 13 72 37 34 334
31 48 17 30 41 20 16 20 56 34 14 88 77 11 12 4 2 4 78 5 18 1108
28 36 28 31 38 19 20 23 51 27 37 35 31 35 8 3 7 4 78 6 25 4198
33 58 28 28 34 18 17 18 40 25 11 84 56 7 10 2 0 5 83 11 33 177
∗ Indicates that a joint t-test of the equality of the means across the reasons was rejected at a 5% significance level.
J.M. Stewart / Journal of Health Economics 20 (2001) 781–796
789
Table 2 presents the composition of the sample by the reason the ROE job ended and by the variables used in the empirical analysis. An asterisk beside a variable name indicates that the joint t-test that the means for each reason for termination are equal was rejected at the 5% significance level. The only variables for which the joint test was not rejected are the proportion of the group that have completed high school and the proportion of the group that have a child less than 6 years old. The reported spell length accounts for censoring by estimating an expected duration for censored spells and assigning the estimated spell length to any censored spells. Individuals who report leaving the ROE job because of illness appear to be older, less educated, and more likely to be female than other categories of individuals. In order to accurately measure the impact of limited health status on the duration of unemployment, it is necessary to extend the analysis of the spell lengths beyond the empirical hazard rate and to control for the impact of these socioeconomic variables. Table 3 compares the socioeconomic composition of individuals who report they have a health limitation and those who report they do not. An asterisk beside a variable name indicates that the difference between the two groups is significantly different from zero at
Table 3 Socioeconomic characteristics of total sample and by presence of a health limitation (enteries are in percentages unless otherwise indicated) Variable
Total
Health limitation
No health limitation
Total sample (%) Age (years)∗ Not married Less than high school∗ High school Female Child under 6 Visible minority Income × 1000 (CAD$)∗ Cohort 2 Full time Union contract No expectation to return No previous UI claim Seasonal job Professional Clerical∗ Service Primary industry Industrial Reason: Ill∗ Quit∗ Laid off∗ Dismissed Spell length∗ (weeks) Sample size
100 36 39 26 31 39 19 19 22 52 72 32 46 41 28 9 3 6 5 77 6 19 72 3 24 5817
8 40 38 36 31 42 19 23 20 47 70 29 38 44 25 7 1 7 6 79 27 12 57 4 35 457
92 36 39 25 31 39 19 19 22 52 72 32 47 41 28 9 3 6 5 77 4 20 73 3 23 5360
∗ Indicates that a joint t-test of the equality of the means across the reasons was rejected at a 5% significance level.
790
J.M. Stewart / Journal of Health Economics 20 (2001) 781–796
a 5% significance level using a t-test. Individuals who report a limitation are older, less educated, have a lower income, are more likely to have left the ROE job because of illness and less likely to have quit or been laid off than those who do not report a limitation. About 27% of individuals with a limitation left their ROE job because of illness or injury compared to only 4% of those without a limitation. 5. Results 5.1. Hazard estimation Table 4 presents the results from four specifications of the hazard function. Two functional forms of the independent variables for health status are used. The first functional form includes interaction terms of the two health status indicators. The interaction terms are included because there are a number of individuals who reported both leaving the ROE job because of illness and having a health limitation. The second functional form includes separate indicators for the presence of a health limitation and for the reason for termination. The omitted category in both functional forms is an individual who does not have a health limitation and who was laid off. Each functional form is used in specifications that do and do not control for unobserved heterogeneity. The coefficients in Table 4 are the exponentials of the actual coefficients estimated and represent the proportional shift in the baseline hazard due to a change in the variable. A value greater than one indicates an upward shift, or higher hazard rate, while a value less than one indicates a downward shift, or lower hazard rate. Having a higher hazard rate indicates that the individual is more likely to leave unemployment and, therefore, expected to have a shorter unemployment spell and vice versa for a lower hazard rate. The estimated coefficients for the explanatory variables are similar across the two forms of incorporating the health status variables, but they do differ, in some cases, depending on whether or not the specification controlled for unobserved heterogeneity. In all four models, the baseline hazard rate refers to the an individual with the same characteristics and so it is not surprising that including the health status variables as interaction terms rather than separate dummy variables does not affect the estimated coefficients for other explanatory variables. Unobserved heterogeneity can be thought of as an omitted variable, therefore, when it is controlled for it would be expected that coefficients would change if they are correlated with the missing information. The estimated coefficients are inline with expectations. For example, older individuals are predicted to have lower hazard rates, or longer unemployment spells. However, when controlling for unobserved heterogeneity, it is only the two older age groups that have statistically significant lower hazard rates, rather than a consistent difference across all age groups. The specifications that control for unobserved heterogeneity indicate that it exists and, therefore, the preferred specifications are those which control for unobserved heterogeneity. The coefficients of the variables for health status are fairly consistent across specifications. In the specifications which use the interaction terms, having a health limitation is associated with a lower hazard rate compared to those without a health limitation and left the ROE job
J.M. Stewart / Journal of Health Economics 20 (2001) 781–796
791
792
J.M. Stewart / Journal of Health Economics 20 (2001) 781–796
J.M. Stewart / Journal of Health Economics 20 (2001) 781–796
793
for the same reason. Many of the coefficients for the interaction terms are not statistically significant, but individuals with a health limitation and left the ROE job because of illness have a statistically significant lower hazard rate and those who quit the ROE job have a statistically significant higher hazard rate. The results are similar when separate dummy variables are used for the health status indicators. Individuals with a health limitation have a statistically significant lower hazard rate. Leaving the ROE job because of illness is also associated with a statistically significant lower hazard rate while quitting the ROE job is associated with a statistically significant higher hazard rate. These results indicate that impaired health status leads to a lower probability of leaving unemployment and, therefore, longer unemployment spells. 5.2. Effect of hazards on composition of stock of the unemployed Some simplifying assumptions must be made in order to examine the impact of longer unemployment spells of individuals with impaired health status on their share of the stock of unemployed. First, assume a steady state, that is, the composition of the flow into unemployment is constant over time. Second, assume that all individuals are homogenous except for their health status which is classified into eight groups; each of the four reasons for a job termination with and without a health limitation. Table 5 shows the proportion of the flow into unemployment, average duration, and implied proportion of the stock of unemployment for each of the eight groups. The top three rows of Table 5 are the groups who are healthy and the bottom five rows have impaired health status. The share of the flow into unemployment for the top three groups combined is 88.54% and for the groups with impaired health status it is 11.46%. When the differences in average duration are taken into account, the shares in the stock of unemployment are 78.55% for the healthy groups and 21.45% for the groups with impaired health status. Although, individuals with lower health status comprised only 11.46% of the flow into unemployment, they account for 21.45% of the stock of unemployed, which is almost double their entrance share, because of their longer unemployment spells. What are the implications for a study that has compared mortality rates of the stock of employed and unemployed? Assume that individuals with high health status have lower mortality rates than those with impaired health status, but that there is no difference in Table 5 Flow composition, average duration, and implied stock composition Group
Flowi
Average duration
Stock i
Laid off no health limitation Quit no health limitation Dismissed no health limitation Illness no health limitation Illness health limitation Dismissed health limitation Quit health limitation Laid off health limitation
67.70 18.14 2.70 3.61 2.13 0.34 0.91 4.47
3.13 0.25 3.53 4.43 12.83 4.54 0.86 3.61
73.68 1.57 3.30 5.56 9.48 0.54 0.27 5.60
794
J.M. Stewart / Journal of Health Economics 20 (2001) 781–796
mortality rates between employed and unemployed individuals. Also, assume that there is no selection bias into the flow into unemployment by health status so that the composition of the stock of employed by health status is the same as the flow out of it. These restrictions imply that the estimate of the impact presented here is a lower bound in some sense. The predicted mortality rate for the stock of employed (MRE ) will be a weighted average of the mortality rate for healthy individuals (MRH ) and the mortality rate for individuals with impaired health (MRIH ) MRE = (0.8854)MRH + (0.1146)MRIH
(15)
where the proportion of the stock in each health status is from the data. The predicted mortality rate for the stock of unemployed (MRU ) is MRU = (0.7855)MRH + (0.2145)MRIH
(16)
where the proportion of the stock in each health status is from the estimated stock share, when the differences in the durations is taken into account. The difference between the mortality rates is MRU −MRE = (−0.0999)MRH + (0.0999)MRIH = (0.0999)(MRIH −MRH )
(17)
Even if unemployment has no impact on health status, the relative mortality rate between the unemployed and the employed will still be greater than one. Can the results in this paper fully explain the difference in mortality rates between the unemployed and employed observed in previous papers? If we examine the lower bound of the estimate from Iversen et al. (1987), which is a relative mortality rate of 1.33, then MRU (0.7855)MRH + (0.2145)MRIH = = 1.33 MRE (0.8854)MRH + (0.1146)MRIH
(18)
which would imply that MRIH = 6.31 MRH
(19)
It would be necessary for the relative mortality rate between individuals with impaired health and those without impaired health to be around 6.3 for the results from this paper to fully explain the lower bound of the previously observed differences. This condition seems unlikely to be met. Those with a health limitation only compose a small part of the flow into unemployment, most individuals enter unemployment because they have been laid off. Even though the longer unemployment spells doubles the proportion of the stock of the unemployed that have a health limitation compared to their proportion of the flow into unemployment, they still only comprise around 20% of the stock of unemployed. In order for this small proportion to have a significant impact on the overall mortality rate of the stock of unemployed, their relative mortality rate to those without a health limitation needs to be high. The simplifying assumptions that have been made must also be kept in mind when considering this result. It has been assumed that the only differences between individuals are the reason they left the ROE job and the presence of a health limitation.
J.M. Stewart / Journal of Health Economics 20 (2001) 781–796
795
The impact of other socioeconomic characteristics which are correlated with limited health status has not been included. Although, health-related selection due to longer unemployment spells does appear to occur and implies that previous research overestimates the impact of unemployment on mortality, it explains only some of the observed difference in mortality rates between the unemployed and employed.
6. Conclusion Numerous studies have found an empirical link between unemployment and mortality rates. One hypothesis is that unemployment decreases health status, but this relationship could be observed for three reasons. First, unemployment does decrease health status. Second, individuals with impaired health status are more likely to become unemployed and, therefore, comprise a larger proportion of the stock of unemployed. Third, individuals with impaired health are more likely to remain unemployed and, therefore, will comprise a larger proportion of the stock of unemployed. This paper focused on the third issue. Are individuals with impaired health more likely to remain unemployed? Duration analysis was used to estimate the effect of health status and other socioeconomic factors on the length of unemployment spells using a longitudinal data set that interviewed individuals after a job separation. Two proxies for impaired health were examined in this paper: a job termination because of illness and a reported health limitation. The PGM procedure was used to estimate a hazard function which allowed for controls for unobserved heterogeneity. A job termination due to illness and a reported health limitation both had a significant, negative impact on the hazard rate. The probability of finding employment at any point in time is lower if the individual quit due to illness or had a health limitation. In particular, individuals who reported both a job termination due to illness and a health limitation had long unemployment spells. The stock of unemployed will be comprised of a larger proportion of individuals with impaired health than the stock of employed because of the longer unemployment spells of these individuals. The implication for previous research is that the estimated impact of unemployment on health status will be an overestimate. To accurately measure the social cost of unemployment the results in this paper, and the results of other researchers, need to be taken into account. References Arrow, J.O., 1996. Estimating the influence of health as a risk factor of unemployment: a survival analysis of employment durations for workers surveyed in the German socioeconomic panel. Social Science and Medicine 42 (12), 1651–1659. Bartley, M., Owen, C., 1996. Relation between socioeconomic status, employment, and health during economic change, 1973–1993. British Medical Journal 313, 445–449. Brenner, M.H., 1979. Mortality and the national economy: a review and the experience of England and Wales, 1936–76. Lancet, September 15, 568–573. Hemstrom, O., 1999. Explaining differential rates of mortality decline for Swedish men and women: a time series analysis, 1945–1992. Social Science and Medicine 48, 1759–1777.
796
J.M. Stewart / Journal of Health Economics 20 (2001) 781–796
Iversen, L., Andersen, O., Andersen, P.K., Christoffersen, K., Keiding, N., 1987. Unemployment and mortality in Denmark, 1970–80. British Medical Journal 295, 878–884. Joyce, T.J., 1989. A time series analysis of unemployment and health: the case of Birth outcomes in New York city. Journal of Health Economics 8 (4), 419–436. Joyce, T.J., Mocan, N., 1993. Unemployment and infant health: time series evidence from the state of Tennessee. Journal of Human Resources 28 (1), 185–203. Junankar, P.N., 1991. Unemployment and mortality in England and Wales: a preliminary analysis. Oxford Economic Papers 43 (2), 305–320. Lancaster, T., 1990. The Econometric Analysis of Transition Data. Cambridge University Press, Cambridge. Magnusson, S., Sigurdsson, J.A., Svardsudd, K., 1999. Indicators of health and well-being in Iceland and Sweden: a comparative study of various indicators concerning standards of living and mortality. Scandinavian Journal of Primary Health Care 17, 139–144. Martikainen, P.T., 1990. Unemployment and mortality among Finnish men, 1981–5. British Medical Journal 301, 407–411. Martikainen, P.T., Valkonen, T., 1996. Excess mortality of unemployed men and women during a period of rapidly increasing unemployment. Lancet 348, 909–912 . Martikainen, P.T., Valkonen, T., 1998. The effects of differential unemployment rate increases of occupation groups on changes in mortality. American Journal of Public Health 88 (12), 1859–1861. McAvinchey, I.D., 1984. Economic factors and mortality: some aspects of the Scottish case, 1950–1978. Scottish Journal of Political Economy 31 (1), 1–27. McAvinchey, I.D., 1988. A comparison of unemployment, income and mortality interaction for five European Countries. Applied Economics 20 (4), 453–471. Meyer, B.D., 1987. Hazard and Markov Chain Models with Applications to Labour Economics. Thesis MIT. Meyer, B.D., 1990. On unemployment insurance and unemployment spells. Econometrica 58 (4), 757–782. Montgomery, S.M., Bartley, M.J., Cook, D.G., Wadsworth, M.E.J., 1996. Health and social precursors of unemployment in young men in Great Britain. Journal of Epidemiology and Community Health 50, 415– 422. Morris, J.K., Cook, D.G., Shaper, A.G., 1994. Loss of employment and mortality. British Medical Journal 308, 1135–1139. Moser, K.A., Fox, A.J., Jones, D.R., 1984. Unemployment and mortality in the OPCS longitudinal study. Lancet, 928–942. Moser, K.A., Fox, A.J., Jones, D.R., Goldblatt, P.O., 1986. Unemployment and mortality: further evidence from the OPCS longitudinal study, 1971–1981. Lancet, 365–366. Moser, K.A., Goldblatt, P.O., Fox, A.J., Jones, D.R., 1987. Unemployment and mortality: comparison of the 1971 and 1981 longitudinal census samples. British Medical Journal 294, 86–90. Prentice, R., Gloeckler, L., 1978. Regression analysis of grouped survival data with an application to breast cancer. Biometrics 34, 57–67. Ruhm, C.J., 1996. Are Recessions Good For Your Health? NBER Working Paper 5570. Stewart, J.M., 1999. Three Essays on Unemployment and Social Assistance. Thesis McMaster University. Wadsworth, M.E.J., Montgomery, S.M., Bartley, M.J., 1999. The persisting effect of unemployment on health and social well-being in men early in working life. Social Science and Medicine 48, 1491–1499. Wagstaff, A., 1985. Time series analysis of the relationship between unemployment and mortality: a survey of econometric techniques and replications of Brenner’s study. Social Science and Medicine 21 (9), 985–996. Yang, B., Lester, D., 1995. Suicide, homicide and unemployment. Applied Economic Letters 2, 278–279.