The effects of depressive symptoms on earnings.
A widespread view of depression is that it puts an enormous social and economic burden on both the individual and society. Depression may reduce the individual's productivity, increase absenteeism and cut-back days, and cause loss of employment (Kessler and Frank 1997; Berndt et al. 1998; Lim, Sanderson, and Andrews 2000; Alexandre and French 2001; Marcotte and Wilcox-Gok 2001; Stewart et al. 2003; Alexandre, Fede, and Mullings 2004). If depression is treated, there are further costs in terms of payments or co-payments related to office visits, costs of medication, and time cost of the treatment. In the United States depression alone had a total cost of more than $83 billion in the year 2000, of which $51.5 billion was the estimated cost at the workplace (Greenberg et al. 2003). Much of this cost is borne by the affected individual.
In this study, I focus on the wage impacts of depression, which is one of the most common diagnoses in the health care sector (World Health Organization 2001), by using the National Longitudinal Survey of Youth 1979 (NLSY79). It has been estimated that the lifetime prevalence of major depressive disorder is 16.2%, and the annual incidence rate is about 6.6% in the United States (Kessler et al. 2003). Antidepressants ranked fourth in annual sales among medications, with $10.2 billion in sales, and they were the fifth most frequently prescribed medications in the United States in 2005 (Intercontinental Marketing Services Health 2006). More than 122 million antidepressant prescriptions were written in 2005, yet depression and other mental disorders are generally underdiagnosed and undertreated (U.S. Department of Health and Human Services 1999; World Health Organization 2001).
The causal effect of depression on workplace performance is difficult to measure, since the relationship between depression and labor market outcomes is rather complex. Depressed people could have lower earnings potential for several reasons. First, and most importantly, depression could hinder efficiency at the workplace. This can take various forms. For instance, during times of depression workers may find it harder to concentrate on a particular task, and everything may seem to take extra effort, or depression could increase the number of missing workdays (absenteeism) or the amount of lost productivity time (Kessler et al. 1999; Stewart et al. 2003). On the other hand, it is possible that poor labor market outcomes or stressful work environments create mental problems (Marcotte, Wilcox-Gok, and Redmon 1999; Clark, Georgellis, and Sanfey 2001; Paterniti et al. 2002). At the same time, having a job can be beneficial to the extent that employment facilitates the creation of social networks; colleagues can help those who are in need of emotional support. Furthermore, people who are prone to periods of mental stress may prefer jobs that provide generous health insurance, or alternatively, these people may select into jobs that can be done even while the individual is in a fragile mental state and that will allow them to be as productive as people without the mental illness.
Finally, it is also possible that depression itself is not the cause or consequence of poor performance, but that it is instead an observable event that occurs in certain people with certain characteristics and that these people tend to perform below average because of the same characteristics that make them prone to psychological imbalance. In other words, depression could merely reflect certain personality traits that are also correlated with productivity.
In the psychology literature, it is well documented that people with certain types of underlying character or personality traits are more prone to depression than are people who lack these traits (Clark, Watson, and Mineka 1994; Hettema 1995; Robins 1995; Watson and Clark 1995; Clark et al. 2003). Personality, on the other hand, has been documented to play an important role in hiring decisions and in later performance in the workplace (Barrick and Mount 1991; Tett, Jackson, and Rothstein 1991; Hogan, Hogan, and Roberts 1996; Goodstein and Lanyon 1999; Bono and Judge 2003). Usual types of personal characteristics that are important in the workplace and for the psychological wellness of the individual include commitment, responsibility, enthusiasm, activity, and optimistic life orientation positive characteristics; and sensitivity, indecisiveness, dependency, passivity, and pessimistic life orientation--negative personal characteristics.
My paper is an innovation in that it takes advantage of a longitudinal survey design that features the same mental health information in multiple years. I use fixed-effects (FE) estimators to eliminate the unobserved heterogeneity that might have biased the results in previous studies. Therefore, by using multiple years of information from a panel survey, I am able to control for personal characteristics that seem to be important when estimating the effects of mental health on productivity. I find that ordinary least-squares (OLS) estimates are negative and strongly significant; however, FE estimates are smaller in magnitude, and they are mostly insignificant. My results indicate that taking personal characteristics into account reduces the magnitude of coefficient estimates of depressive symptoms.
Figure 1 foretells the main story of the paper. The four lines in the figure represent the wage profile of four groups of people in the NLSY79. The thin dashed line represents the wage evolution of those who do not show depressive symptoms in any of the years when NLSY79 administered the Center of Epidemiologic Studies Depression Scale (CES-D) questionnaire. (1) The thin solid line represents the wage evolution of those who show depressive symptoms in both the early period (either 1992 or 1994) and the late period. (2) The thick dashed line represents the wage evolvement of those who show depressive symptoms in the early years but no symptoms in the late period, and the thick solid line represents those who show depressive symptoms in the late period but do not have symptoms in the early years.
[FIGURE 1 OMITTED]
The graph depicts simple means of logarithm of hourly wages throughout time without controlling for other potential covariates that affect wages. However, this graph is rather indicative of the results of this study. By looking at the dynamics of the four lines, we would expect OLS models to produce negative mental health coefficients on the depression dummy in a cross-sectional analysis. Figure 1 also indicates that a cross-sectional analysis would not give insights to the full story. The graph reveals the importance of an analysis that takes time into consideration and pays attention to the dynamics of the evolution of wages. From Figure 1, it is easily noticeable that there are three distinctly different groups of people in terms of depressive symptoms. (3) It seems that people who never show depressive signs are inherently different from those who have symptoms more than once during the analyzed period. Additionally, there are the symptom changers, who seem to be remarkably similar to each other in terms of their wage development and inherently different from both the no-symptom group and those who show depressive symptoms in both periods. If mental distortions disrupt productivity, a separate drop would be expected to occur around the 1992-1994 period for those who show symptoms in the early years but not in the later period. Another drop would be expected to occur around the late years for those who did not show symptoms in the early period but did in the late years. Contrary to expectations, the two lines representing the symptom changers line up very nicely. It is possible that these two groups in the middle are not much different overall; they could be prone to depression, and the only difference between them is the timing of the depressive spell. The timing of their depressive shocks could be random; they come out of and go into depression spells at certain times, but their productivity seems to be unaffected by the timing of these shocks. This could occur because these two groups may have very similar characteristics that signal their levels of productivity to their employers.
The remainder of the paper is arranged as follows: In section 2, I review the findings of the existing literature, and in section 3, I introduce the econometric tools that I apply. Section 4 describes the data, section 5 lays out the main results of the paper, and concluding remarks can be found in section 6.
2. Literature Review
Previous studies in economics have mainly focused on the wage effects of psychological problems in general. In the early literature, mental health is mostly assumed to be exogenous. Bartel and Taubman (1979) use the National Academy of Sciences--National Research Council twin sample and estimate that people with psychoses/neuroses have 16-27% lower earnings. In another paper, they revisit the problem of mental illness and find that psychoses are more impeding: They reduce wages by 32-47%, and neuroses cause a 12-14% reduction (Bartel and Taubman 1986). Frank and Gertler (1991) criticize some of the early studies because of their reliance on health care utilization without information of current mental health status, which could lead to substantially biased estimates. Their analysis, which uses mental health measures generated from survey responses, shows that men with mental disorder have about 21% lower earnings. Recognizing the problem of potential simultaneity, Ettner, Frank, and Kessler (1997) use Instrumental Variables (IV) methods to obtain estimates of the genuine effect of mental illness using the National Comorbidity Survey (NCS). The instruments used in the analysis are the number of psychiatric disorders exhibited by the respondent's parents and the number of psychiatric disorders experienced before age 18 by the respondent. For women, they find a 28% decrease in annual income and a 14% reduction in employment if any disorder is diagnosed. For men, diagnosis reduces employment by 12%, and they, too, have lower wages (by 9.5%), but the effect on wages is insignificant. Interestingly, their IV results are not significantly different than the single-equation results in any of the specifications. The disadvantage of using lagged mental health variables as instruments is that if both current and previous mental health are correlated with personality, as shown by the psychology literature, the instruments could be correlated with the dependent variable not only through current mental illness but through personal characteristics as well, since they have also been found to be important in work performance.
Baldwin and Marcus (2006) use the Medical Expenditure Panel Survey to investigate whether discrimination (stigma) could be a reason that mentally ill people have worse labor market outcomes. They find that people with mood disorders have about 5.6-9.5% lower predicted wages, and that 88% of this is explained by differences in other explanatory variables. Their analysis does not discuss the potential problems of endogeneity that may be present in the wage equations. In another recent study, Jofre-Bonet et al. (2005) use the Community Tracking Survey to analyze the impact of poor mental health and smoking. Their most basic specification finds that poor mental health reduces men's hourly wages by 7.8% and women's by 4.2%. By interacting poor mental health and smoking, they also find that current male smokers in poor mental health earn 16.3% lower wages, while the reduction for current female smokers in poor mental health is 5.3%.
Although the number of studies analyzing the labor market effects of mental health is growing at a rapid rate, research that singles out depression as the primary focus of analysis is rather sparse. My paper is closest in subject to that of Marcotte, Wilcox-Gok, and Redmon (2000), who, using the NCS, analyze the labor market effects of affective disorders: major depression, dysthymia (a longer-term depressed mood that does not qualify for major depression), and bipolar disorders. They find that none of these three significantly impacts the productivity of men; however, their results indicate that the yearly income of women with depression is about $6000 less than that of women without depression (and this estimate is significant at the 5% level). Since the mean yearly income is around $18,500 for women in their sample, this loss is equivalent to about a 32% reduction in wages. Interestingly, the coefficient of dysthymia is positive but insignificant for women. They use history of affective disorders of the parents as the instrument. (4)
While some of the above studies use longitudinal surveys, none of them uses panel data techniques to estimate the effects of mental health. To my knowledge, this is the first paper to estimate the wage effects of mental health (in particular, depressive symptoms) by taking advantage of the panel nature of the NLSY79.
If we can assume that depression is exogenous and randomly distributed in the population, OLS methods can be applied to estimate the coefficients of a simple wage equation such as the following:
ln [w.sub.it] = [X.sub.it][beta] + [delta][D.sub.it] + [[kappa].sub.t] + [u.sub.it], (1)
where ln [w.sub.it] is the logarithm of hourly wages for person i in year t. [X.sub.it] represents usual explanatory variables (age, marital status, race, education, occupation and industry dummies, etc.); [D.sub.it] is a measure of depression; [[kappa].sub.t] is a set of time dummies; and [u.sub.it] is an error term. The estimated coefficients from the above model are consistent and unbiased if the error term has the assumed nice properties, most importantly that it is uncorrelated with the explanatory variables.
If, on the other hand, the true model should include time-invariant personal characteristics, Equation 1 has to be extended to
ln [w.sub.it] = [X.sub.it][beta] + [delta][D.sub.it] + [[kappa].sub.t] + [chi][a.sub.i] + [u.sub.it], (2)
where [a.sub.i] represents unobserved (time-invariant) personal characteristics. If Equation 2 is the true model, OLS estimates of Equation 1 suffer from omitted variables bias if Corr([a.sub.i], ln [w.sub.it]) [not equal to] 0 and Corr([a.sub.i], [D.sub.it]) [not equal to] 0 (Wooldridge 2002). In terms of the present problem, [a.sub.i] could be thought of as some personality traits that are correlated with both productivity and depression. Based on the psychology and management literature mentioned above, this is a plausible case.
Longitudinal surveys allow us to eliminate the problem of unobservable heterogeneity under the assumption that the heterogeneity is time invariant. In a frequently cited meta-analysis, Conley (1984) concludes that personality traits show remarkable longitudinal stability even many years apart. (5) Gustavsson et al. (1997) also find personality traits to be stable over the nine-year period of their analysis. Furthermore, they also find that personality traits are significant predictors of emotional and physical health as well as marital functioning and job stress. According to McCrae and Costa (1994), personality traits reach stability at around age 30, and after age 30 there are only subtle changes in personality that are mostly consistent with a lower level of activity. (6) Therefore, I can assume that the application of a FE estimator is reasonable to reduce the contamination caused by unobserved personality dimensions.
The standard FE estimators are the first-difference estimator and the mean-difference estimator,
[DELTA] ln [w.sub.it] = [DELTA][X.sub.it][beta] + [DELTA][D.sub.it] x [delta] + [[??].sub.t] + [DELTA][u.sub.it], (3)
where [DELTA] indicates first differencing or mean differencing, [[??].sub.t] represents change in time, and where personality traits are not in the model, considering that the unobservable heterogeneity is truly time invariant ([a.sub.it] = [a.sub.it-1]). Assuming that the idiosyncratic error is uncorrelated with the changes in the explanatory variables, using OLS for Equation 3 will provide consistent estimates.
The first-difference and the mean-difference estimators yield the same results for two periods; however, they can differ when more periods are considered. A general problem with first-difference models is that they can inflate measurement errors compared to mean-difference models. Griliches and Hausman (1986) show, however, that when first differencing is done by using years that are further apart, the bias caused by measurement errors is mitigated, compared to the mean-difference estimator. I use both mean-difference and first-difference estimators in the analysis and show that the main conclusion does not depend on which method is applied. I also attempt to tackle the problem of measurement error bias by using both a dummy variable and an index variable as measures of depression. The main findings are robust to the choice of the type of the explanatory variable as well.
Results of the OLS models could be biased as a result of non-random selection into the working sample. Linear regression results reported in this paper are corrected for this bias by utilizing the Heckman sample selection model. Variables included in the first stage--and not included in the wage equations--are number of children below six years of age, number of children between the ages of 6 and 18 years, and non-labor income. In estimating the FE results, I take selection into account by following the method of Mundlak (1978). First a probit model on employment is estimated for a pooled sample, in which the explanatory variables include all the right-hand-side variables from Equation 1 along with each individual's time-mean values. This model is used to construct the inverse Mills ratio, which is then included in the FE model.
Alternatively, IV techniques could be applied to address the problem of simultaneity. If a set of instruments can be found that is correlated with depressive symptoms but not correlated with wages, except through depression, a 2SLS estimator can be applied. After exploring several sets of potential instruments to predict depressive symptoms, such as the death of a child, death of the other parent of a child, happiness in marriage for married women, and in 2004 the presence of thyroid problems and change in menstrual patterns for women, I obtained first stages with small coefficients of determination ([R.sup.2] values), mostly below 0.12, and partial [R.sup.2] values below 0.04. (7) Consequently, IV techniques generated a fivefold to tenfold increase in standard errors compared to OLS results.
Even if these instruments worked well in the model to predict depressive symptoms, one could still argue that they are not valid. The main concern with life events such as divorce/ separation, financial problems, serious housing problems, serious illness or injury, job loss, legal problems, death in the family, etc., is that they are likely to be correlated with wages. Furthermore, major life events that most often have the strongest impact on depressive symptoms are rather rare and infrequent. Minor life events, on the other hand, may not contribute to the likelihood of depression (Brown and Harris 1986). Moreover, life events may not be totally random. Individuals with certain characteristics (ones that are also associated with likelihood of depression) are more likely to select into situations with a high probability of producing stressful life events (Kendler, Karkowski, and Prescott 1999). Also, perceptions of stress can result from psychopathology. Individuals in a depressive state are more likely to recall minor incidents as more stressful major events (Monroe and Simons 1991).
The problem of weak instruments is common (see, for instance, Levine, Gustafson, and Velenchik 1997; Bray 2005). Weak instruments can cause serious bias that many times makes IV estimates worse than OLS estimates (Bound, Jaeger, and Baker 1995). To the extent that the panel data techniques used in this paper cannot eliminate the bias from simultaneity, the FE results are biased. Low wages could increase stress through the financial uncertainty they cause some people or through lower amounts of health investments. To the extent that the hypothesized and documented impacts of depression include lower wages, the simultaneity would inflate the impact of depression on wages in all specifications (the FE model included). This means that the true coefficient of depressive symptoms should be less negative than what is estimated below. (8)
My analysis uses data from the NLSY79, an ongoing study that initially provided information for a sample of 12,686 people between the ages of 14 and 22 years. Until 1994, the NLSY was conducted each year; since then it has been a biennial survey. The primary purpose of this survey was the collection of data on labor force experiences, investment in education, and training, but from year to year additional topics were also addressed. I am taking advantage of the mental health--related questions to investigate how depressive symptoms affect labor market outcomes.
The NLSY79 asks some of the CES-D questions. The CES-D is one of the most frequently used depression questionnaires that psychologists developed to detect the presence of depressive symptoms in general surveys. (9) The original CES-D scale comprises a 20-item questionnaire that asks questions about emotions felt recently by the respondents. With four response categories (0 = none to 3 = almost every day), the maximum score of the scale is 60, and the presence of depressive symptoms is detected if the individual has a score of 16 or higher. (10) Although the CES-D was not intended as a clinical diagnostic tool, the scale has very good properties that make it favorable for use in household surveys (Radloff 1977). The NLSY79 used the full set of 20 questions in 1992 and then a subset of seven questions in 1994 and as a part of the 'Health Module 40 & Over.' (11,12) Since only seven questions out of the 20 are available in each year, my measures are based on a short CES-D scale. This exact same set of seven questions was used by Ross and Mirowsky (1989), but other shorter versions of the scale have also been used in other studies and have been proved to be similarly reliable compared with the full version (Melchior et al. 1993; Rouch-Leroyer et al. 2000; Grzywacz et al. 2006). After summing the answers, scores of the seven-item scale range from 0 to 21. I use two measures for the main variable of interest: a simple dummy variable and raw scores (ranging from 0 to 21) and its square term. The dummy variable approach is closer to the world of psychiatric diagnosis, in which a person either has or does not have major depressive disorder. The disadvantage of using a dummy, however, is that it cannot capture the severity of the disorder. Using an index variable--the score--has the advantage that it does not collapse the variation in severity of distress into two categories and therefore, preserves more of the information that is provided by the respondent. Furthermore, using a squared term of the score allows for nonlinearity to take place in the model. (13)
To detect the presence of depressive symptoms and to create the dummy variable, I use a cutoff value of seven. (14) I also experimented using other thresholds, but those did not change the main findings of the paper. This scale, using the above threshold, detects the presence of depressive symptoms for 22.32% of the full sample answering this question in 1992, 20.37% in 1994, and 17.6% in 2004. There is a concern that people with (severe) depressive symptoms may become missing or non-respondent in later years. This would mean that my coefficients underestimate the magnitude of the true results because I only detect wage responses of those with symptoms that are not too severe. I therefore estimate probit models to see whether those with symptoms in 1992 or in 1994 are significantly more likely to be missing in later years or are more likely to have missing information if they are interviewed. Results (not shown) indicate that respondents with depressive symptoms are not significantly more likely to have missing information in later years (either as a result of dropping out or as a result of not responding to certain questions).
The longitudinal nature of the NLSY79 allows the analysis of development of wages for those who never report the presence of depressive symptoms and also allows for comparing these individuals to those who change depressive symptoms and those who are depressed in both the earlier years and in the later years as well.
My analysis drops respondents from the sample if they do not report or if they refuse to report the main variables such as wage, CES-D, limitation, and non-labor income for the previous year. (15) Therefore, my samples consist of 3660 male and 3768 female respondents in 1992 (above 80% of the 4481 males and 4535 females interviewed), 3427 males and 3556 females in 1994 (about 78-79% of the 4409 males and 4480 females interviewed), and 3143 males and 3285 females in 2004 (about 81 82% of the 3825 males and 4021 females interviewed for the Health Module questions). The selection equations for the first-difference models are based on a sample of individuals with information for both years used in the models. Second-stage estimates are based on differences in wages for those with wage reported in both years of the analysis. In the case of the mean-difference model, individuals are included in the sample if they have observations in at least two out of the three years used. (16)
Table 1 presents the descriptive statistics in the years 1992, 1994, and 2004 for men and women separately. As shown in the summary table, depressive symptoms are more prevalent among women than men. Depending on the year 13.2-17.4% of the male sample have depressive symptoms, while the same numbers for females are 20.6-26%. Males are more likely to be employed, and on average they command higher wages. In 1992 about 8.9% of the male sample have certain work limitations, and this percentage grows to about 10% when they reach the age of 40. About 5% of the female sample have work limitations in 1992, and this percentage grows to above 12.8% by the 'Health Module 40 & Over' years. (17) Comparing men and women with and without depressive symptoms (not shown) reveals that people with the symptoms are more likely to report some sort of limitation in the amount or kind of work. Minorities are overrepresented in the sample.
Table 2 shows the main coefficients of interest from simple linear regressions after correcting for selection (18) The table shows two separate specifications for both men and women in different years and in a pooled sample. The first specification (part A for both men and women) uses a single dummy indicating the presence of depressive symptoms. These results indicate that although more prevalent among women, depressive symptoms seem to impede men's productivity more. The coefficient estimates indicate that men with depressive symptoms earn, on average, between 11.7-14.6% less than men without the symptoms. (19) On the other hand, women suffer a loss of between 2% and 5.5%; however, the coefficient estimates using only 1992 or 2004 are insignificant for women. (20) The second specification (part B) shows results using the CES-D score and its squared term. Using the raw score and its squared term is useful for testing whether any nonlinearity is present. People with more serious forms of mental illness could be relatively more disabled, in which case there would be a negative coefficient estimate for both the raw score and its square. On the other hand, a positive coefficient on the squared term might mean that beyond a certain level of severity in symptoms further decline in mental health has little effect on wages. Depression in this latter case would have a deteriorating impact but would be less and less impeding. Table 2 reveals that the squared term is mostly positive (with the exception of 1992 in the male sample); however, none of the squared coefficients is significant, and therefore, there is no strong evidence for nonlinearity in the impact of depressive symptoms. Coefficient estimates of the raw score and the squared term, however, with the exception of 2004 for the female sample, are jointly significant. The main problem with using the score and its square term is that it is rather difficult to interpret what each point increase means on such a scale. It is perhaps not too much of a stretch to imagine that higher scores represent more serious distress, though. The estimated coefficients can be translated as meaning that men with two standard deviations above the average score have on average 8.8-12.2% lower wages, while women two standard deviations above the average score earn 2.6-3.9% less, on average. (21)
These linear regression results, however, grant too much importance to depression if other variables that are positively correlated with depression and negatively correlated with productivity are omitted from the model. For instance, if underlying personal characteristics make people both more prone to develop depressive symptoms and less productive in the workplace, these estimated coefficients may be too large in magnitude. Next, I look for some evidence that depression could reflect some underlying characteristics of these people.
Table 3 exhibits further OLS estimates from wage regressions. The first row of the table contains the average of the raw scores from the three years during which respondents answered the CES-D questions in the NLSY79 (1992, 1994, and 2004). This average score is supposed to capture the respondent's underlying psychological well-being. The second row contains the differences from the average score in each year. To clarify this through an example, consider that there is a respondent with a score of 4 in 1992, a score of 9 in 1994, and a score of 5 in 2004. This person has an average score of 6, which will be used in all three years. The difference from the average will vary each year, though. It will be -2 in 1992, 3 in 1994, and -1 in 2004. It could be that going into a mentally imbalanced spell hampers productivity; that is, when people have worse mental health than their average, their productivity is greatly impacted. If it is the onset of these worse-than-average spells that makes these people less productive, we expect the difference from the average score to be negative and significant. Table 3, in contrast, demonstrates that the average is negative and significant but that the difference from the average is not. These results indicate that if there are people who change in terms of the severity of their depressive symptoms, they will not be necessarily more productive when they are better than their "average self," and they will not be necessarily less productive when they are in a worse state than their "average self." It is, rather, their average score that counts. This perhaps happens because this average score is more representative of underlying personal characteristics.
To find further evidence for the importance of underlying personality, I also estimated 1992 and 1994 wage equations, but instead of using the symptom dummy from the same year, I included a future symptom dummy from 2004 on restricted samples of men and women without symptoms in 1992 and 1994, respectively (results not shown). The idea is that if personal characteristics do not play any role in who is more likely to develop depression, then a future depressive spell should not be detected in early wage regressions if in those particular years the respondent does not show symptoms. The estimated coefficients, however, indicate that future depressive symptoms already reduce wages in earlier years, even for those who do not report depressive symptoms in the particular years in question. (22) These results can be interpreted as meaning that there are people who experience substantial changes in their mental health and that these people are more likely to be less productive even when they are in a relatively healthy mental state.
In Table 4, I also take advantage of the availability of two measures related to personality in the NLSY79. In the initial year of the survey, a short version of the Rotter Internal-External Locus of Control Scale was administered, and in 1980 and in 1987 respondents answered questions from the Rosenberg Self-Esteem Scale (for further information about these measures see Rosenberg 1965; Rotter 1966; Goldsmith, Veum, and Darity 1997). These measures focus on certain aspects of personality and obviously do not cover the full dimension of underlying personal characteristics; however, their relevance in this study is that both of these measures contribute to vulnerability to depression (Pearlin et al. 1981; Costello 1982). In Table 4, I create four groups similar to the groups in Figure 1 to test whether the respondents' scores on these scales are significantly different. (23) The scores themselves are hard to interpret, but the means of those with symptoms in the early but not the late period (column 2) and those with symptoms in the late but not the early period (column 3) are always between the averages of those without any symptoms and those with symptoms in both early and late periods. More importantly, means in the two middle columns are not significantly different from each other but are always significantly different from means of those in the other columns, at least at the 10% level of significance. This table is just one more piece in the puzzle pointing toward the importance of personality.
Table 5 shows the means of changes in certain variables for four groups between 2004 and 1992. (24) The group of males that shows signs of depressive symptoms in 2004 but not in 1992 seems to have slower wage growth than that of the group comprising those with no symptoms. Those with depressive symptoms in 1992 and no symptoms in 2004 have a similar change in wages compared with those without any symptoms, and their change is not significantly different from that of those with symptoms only in 2004. For females, the largest growth in wages is observed by the group who shows depressive symptoms in 1992 but not in 2004. The group with symptoms in 2004 and no symptoms in 1992 has almost identical growth in wages compared with the one without any symptoms in these years. For females, none of the changes in wages is statistically significant compared with the others. There is one apparent difference for both males and females in the table. People showing signs of depressive symptoms in the later period are also more likely to develop work-limiting conditions than are those in any other group, while the opposite is true for those coming out of a depressive spell. The test of means of changes, of course, does not control for any of the other observables that may affect wages.
To the extent that individual characteristics are time invariant, the FE method takes away the bias caused by not controlling for individual heterogeneity in the cross-sectional models. Table 6 presents these FE estimates in various models. Columns 1 and 2 include long difference results (2004-1992), columns 3 and 4 show the short difference results (1994-1992), and columns 5 and 6 show the mean-difference results including all waves when CES-D questions were administered. Odd-numbered columns in Table 6 contain results with all explanatory variables included along with the inverse Mills ratio to control for selection. Even-numbered columns contain FE estimates from a simple model that only includes the main variable of interest along with year dummies. Estimates in columns 2, 4, and 6, therefore, exclude all variables through which changes in depressive symptoms could affect the outcome (such as change in limitation, marital status, occupation, etc.). In this sense coefficient estimates in the even-numbered columns serve as a sensitivity check to the results and can be considered as a lower bound of the real effect of depressive symptoms.
First, considering the coefficient estimates from the model, including all explanatory variables (columns 1, 3, and 5), it seems that FE estimates are considerably lower than estimates from the cross-sectional models. The only significant coefficient of depressive symptoms for men in the mean-difference model indicates that symptoms decrease wages by about 3.2%. The 95% confidence interval of this estimate does not contain any of the cross-sectional estimates. The coefficient estimates of the dummy in the FE models for men are identified through a sample of 417 symptom changers in the long-difference model, 451 changers in the short-difference model, and 703 symptom changers in the mean-difference model. Coefficient estimates of the CES-D score and its squared term for males are also jointly significant at the 10% level in the short-difference equation (column 3) and in the mean-difference equation (column 5). These coefficients indicate that having a score that is two standard deviations above the average results in a 3.3-4.3% wage loss, which is also lower than what is indicated by the cross-sectional estimates. For females, the only significant estimate indicates a moderate 3.2% decrease in wages if depressive symptoms are present. (25)
Models not controlling for other covariates (in the even-numbered columns) still do not provide evidence for strong negative effects on wages. For men, coefficient estimates of the dummy are significant in the short-difference model (column 4) and in the mean-difference model (column 6). According to these estimates depressive symptoms reduce wages by about 3.4-4.3%. The score and its squared term are also significant in column 4 (at the 1% level) and in column 6 (at the 5% level). These estimates imply a 4-6% reduction in wages for males with two standard deviations above the average. Coefficient estimates for females hardly change when the other control variables are excluded from the models. Estimates for women do not provide evidence for a negative impact on wages (the score and its squared term are not jointly significant in any case for women). These estimates, not controlling for changes in other variables that could lead to changes in wages, however, are likely to be lower bounds for the true effects. Everything that could be correlated with changes in depressive symptoms is now included in the error terms, and coefficients on the main variable are probably biased.
The lower, and mostly insignificant, coefficient estimates of the depression measures indicate that linear regression results in a cross-sectional framework are negatively biased. The results show that taking unobservable (to the researcher) personal characteristics out of the model reduces the effect of depressive symptoms. After all, the results can be interpreted by weaving together what the psychology and the managerial literature showed separately: Personality matters for both depression and productivity. It could be that poorly performing people are more likely to be depressed as well and that there is a common denominator for all these, which is related to the underlying personalities of these individuals.
It is possible, however, that the main effect of depression is not on earnings but rather on labor force participation. For instance, testing differences of the mean CES-D scores in the working and non-working sample reveal that the non-working group always has a significantly higher average score than does the working group. (26) In addition, coefficient estimates of both the depression dummy and the CES-D score and its square term are mostly significant (or jointly significant) in the selection equations, which could also indicate that depression has an important effect on labor force participation. Estimates of the symptom dummy (not all shown) for men indicate a reduction in probability of work by 1.1-5.1%. (27) The same reduction for women is in the 1.6-3.5% range. Similarly, a CES-D score that is two standard deviations above the mean leads to about 1.6-4% reduction in the probability of work for men and a 2.9-4% reduction for women (estimates not shown). Just like the cross-sectional estimates in the wage equation, these estimates could be biased as a result of omitted variables (such as unobserved personality) and/or simultaneity. Results in the first-stage selection equations for the FE models (columns 3 and 4 in Table A1, for instance) indicate that the time-mean of the depressive symptom dummy has a significant effect on the probability of work, but current symptoms do not. (28) This bears some similarity to my results in earnings equations, as for earnings it is the average score of CES-D that seems to matter (see Table 3) and not the difference from the mean.
Next, I estimated FE models to see whether changes in scores or in depressive symptoms explain changes in labor force participation (results not shown). None of the dummy coefficients was found to be significantly different from zero. For males, the coefficients indicate a 0.8-1.8% decrease in the probability of labor force participation. For females, the same effect is between 0% and (positive!) 0.8%. Using the score and its squared term leads to jointly significant estimates in the long-difference model for males and in the short- and mean-difference models for females. These coefficients indicate that someone with two standard deviations above the average score has a 2.67% lower probability of labor force participation for males and a 2.27-2.69% lower probability for females. These FE estimates are also lower than their cross-sectional counterparts; however, the difference between the cross-sectional estimates and the FE results is not as distinctive as in the earnings equations.
6. Concluding Remarks
There is a growing body of literature that estimates the labor market effects of psychological well-being. The number of articles investigating the impact of depression specifically is much smaller. This study is an addition to this latter group. Exploiting the fact that the NLSY79 provides information about depressive symptoms in multiple years, I am able to study how changes in depressive symptoms impact productivity.
Psychological distress variables in cross-sectional models reduce earnings. OLS estimates, however, are likely to be negatively biased as a result of omitted variables. The use of panel data techniques to control for unobserved personal characteristics of the people in my sample reduces coefficient estimates. Taking personal characteristics into account, therefore, proves to be very important, just as in the case of estimating the return to marriage in the case of young men (Cornwell and Rupert 1997) or when estimating the effects of teenage childbearing (Hotz, McElroy, and Sanders 2005).
The OLS estimates in this study are comparable to findings in other studies using logarithm of hourly wages as the dependent variable. My estimates in the OLS models are in the range of 11.7-14.6% decrease in wages for males and 2-5.5% reduction for females. The estimates of Baldwin and Marcus (2006) for the effects of mood disorders on a sample that does not separate men and women indicate a 5.6-9.5% reduction in wages. Jofre-Bonet et al. (2005), in their simplest specification, find that poor mental health reduces men's wages by 7.8% and women's wages by 4.2%. Therefore, their estimates are also similar to mine in that they, too, find stronger effects for men than for women.
My study has several weaknesses that could also prove important in terms of conclusions and inference. The first weakness comes from the nature of the NLSY79, in that it only follows a restricted cohort of people who were 14 21 years old in 1979. At the time of the first wave of the CES-D questions (1992), these people were 27-34 years old. In this respect, the results may not be generalizable to older populations or to individuals who are in their early 20s. Furthermore, to the extent that people with the most serious forms of depression are hospitalized for a longer period, the results could understate the importance of the disease. In this respect, my results can only be generalized to those whose condition does not necessitate long-term hospitalization.
The NLSY79 is quite limited in allowing investigation of how depression affects wages when it is comorbid with other mental illnesses. Some information about illicit drug consumption and alcohol consumption is available; however, this information is not consistently present in the waves of my analysis. Including a variable for any drug use or alcohol consumption and interacting it with depressive symptoms in the years during which data are available does not provide strong evidence that not controlling for alcohol or substance abuse causes bias in the estimates of depression measures. Comorbidity with other mental illnesses, although important, cannot be analyzed with the NLSY79 because of a lack of information about other mental illnesses.
The FE estimators applied in this paper cannot eliminate the bias potential that simultaneity causes. Unfortunately, IV techniques were found to provide weak first-stage (and second-stage) results and therefore to be inadequate to address this problem. As mentioned in the paper, though, to the extent that poor labor market outcomes, such as low wages, are likely to cause depression, my estimates are negatively biased.
Another weakness of this study comes from a lack of complete information regarding depressive symptoms. The CES-D questionnaire is administered at only three points in time. For instance, there is no information about depressive symptoms before 1992, and in the late years of the survey respondents answered the CES-D questions only once depending on availability and on when they reached 40 years of age. This means that it is not possible to apply a full-scale psychological analysis. However, information obtained in those three particular years provides convincing evidence that a cross-sectional analysis is not exhaustive.
As Marcotte, Wilcox-Gok, and Redmon (2000) also argue, "... reliable estimates of the full costs of mental illnesses are vital for policy-makers and health care professionals who seek to properly allocate resources and develop strategies to mitigate the negative consequences of mental illness." The FE results indicate that depression may not be the root of the problem; it may be only one of the symptoms of a larger issue. It could be that the underlying "problems" are personality characteristics, and it is personality that needs to be coached, trained, or improved. And if it is a trait that can be influenced by the environment--that is, if personality is not purely genetic--then perhaps earlier intervention, through which the negative impacts of personality could be attenuated, is preferred. If personality is manageable with early intervention a preferable outcome may be reached (i) by preventing depressive episodes from occurring (or providing the management of the severity of depressive episodes) and (ii) by improving labor market outcomes through better personality characteristics.
[FIGURE A1 OMITTED]
Table A1 Probit Estimates from the Selection Equations with Marginal Effects (Marg. Eff.) Coefficient Estimates from the First-Stage Equations When Second Stage Is an OLS--Pooled Years (2004, 1994, and 1992) Males (1) Marg. Eff. Depressive symptoms -0.257 (0.052) *** -0.039 Children between 0 and 5 0.025 (0.038) 0.003 Children between 6 and 18 -0.004 (0.024) 0.000 Non-labor income (*[10.sup.6]) -0.714 (0.175) *** -0.096 Age 0.004 (0.011) 0.001 Black -0.251 (0.052) *** -0.037 Hispanic 0.021 (0.065) 0.003 Married 0.419 (0.055) *** 0.060 Divorced 0.229 (0.071) *** 0.027 High school 0.324 (0.058) *** 0.047 Bachelor 0.605 (0.088) *** 0.060 More than bachelor 0.467 (0.120) *** 0.046 Urban 0.063 (0.054) 0.009 Limited -1.209 (0.063) *** -0.302 Time-mean of depressive symptoms Time-mean of children between 0 and 5 Time-mean of children between 6 and 18 Time-mean of non-labor income (* [10.sup.6]) Time-mean of age (*100) Time-mean of married Time-mean of divorced Time-mean of high school Time-mean of bachelor Time-mean of more than bachelor Time-mean of urban Time-mean of limited Coefficient Estimates from the First-Stage Equations When Second Stage Is an OLS--Pooled Years (2004, 1994, and 1992) Females (2) Marg. Eff. Depressive symptoms -0.105 (0.037) *** -0.027 Children between 0 and 5 -0.462 (0.027) *** -0.116 Children between 6 and 18 -0.151 (0.016) *** -0.038 Non-labor income (*[10.sup.6]) -0.618 (0.141) *** -0.156 Age 0.002 (0.009) 0.001 Black 0.037 (0.047) 0.009 Hispanic 0.069 (0.054) 0.017 Married 0.256 (0.043) *** 0.065 Divorced 0.374 (0.056) *** 0.082 High school 0.583 (0.053) *** 0.161 Bachelor 0.788 (0.072) *** 0.149 More than bachelor 1.016 (0.109) *** 0.154 Urban 0.009 (0.043) 0.002 Limited -0.823 (0.049) *** -0.265 Time-mean of depressive symptoms Time-mean of children between 0 and 5 Time-mean of children between 6 and 18 Time-mean of non-labor income (* [10.sup.6]) Time-mean of age (*100) Time-mean of married Time-mean of divorced Time-mean of high school Time-mean of bachelor Time-mean of more than bachelor Time-mean of urban Time-mean of limited Coefficient Estimate from the First-Stage Equations When Second Stage Is the Mean-Difference Model Males (3) Marg. Eff. Depressive symptoms -0.073 (0.078) -0.009 Children between 0 and 5 0.129 (0.062) ** 0.015 Children between 6 and 18 0.046 (0.044) 0.005 Non-labor income (*[10.sup.6]) -0.503 (0.253) ** -0.059 Age 0.107 (0.156) 0.013 Black -0.149 (0.055) *** -0.018 Hispanic 0.037 (0.063) 0.004 Married 0.116 (0.122) 0.014 Divorced 0.169 (0.147) 0.018 High school -0.572 (0.249) ** -0.059 Bachelor -0.439 (0.437) -0.065 More than bachelor -0.799 (0.590) -0.154 Urban 0.116 (0.102) 0.014 Limited -0.399 (0.104) *** -0.060 Time-mean of depressive -0.319 (0.110) *** -0.037 symptoms Time-mean of children -0.152 (0.080) * -0.018 between 0 and 5 Time-mean of children -0.095 (0.053) * -0.011 between 6 and 18 Time-mean of non-labor -0.934 (0.423) ** -0.109 income (* [10.sup.6]) Time-mean of age (*100) -0.099 (0.157) -0.012 Time-mean of married 0.451 (0.141) *** 0.053 Time-mean of divorced 0.007 (0.167) 0.001 Time-mean of high school 0.881 (0.254) *** 0.103 Time-mean of bachelor 1.004 (0.444) ** 0.117 Time-mean of more than 1.195 (0.602) ** 0.139 bachelor Time-mean of urban -0.134 (0.124) -0.016 Time-mean of limited -1.302 (0.131) *** -0.152 Coefficient Estimate from the First-Stage Equations When Second Stage Is the Mean-Difference Model Females (4) Marg. Eff. Depressive symptoms 0.002 (0.054) 0.000 Children between 0 and 5 -0.372 (0.045) *** -0.091 Children between 6 and 18 -0.118 (0.031) *** -0.029 Non-labor income (*[10.sup.6]) 0.133 (0.204) 0.033 Age 0.001 (0.131) 0.000 Black 0.065 (0.044) 0.016 Hispanic 0.096 (0.047) ** 0.023 Married -0.035 (0.088) -0.009 Divorced 0.018 (0.108) 0.004 High school 0.210 (0.178) 0.053 Bachelor 0.380 (0.282) 0.081 More than bachelor 0.293 (0.361) 0.063 Urban -0.020 (0.071) -0.005 Limited -0.378 (0.074) *** -0.106 Time-mean of depressive -0.179 (0.075) ** -0.044 symptoms Time-mean of children -0.123 (0.056) ** -0.030 between 0 and 5 Time-mean of children -0.042 (0.035) -0.010 between 6 and 18 Time-mean of non-labor -1.550 (0.292) *** -0.381 income (* [10.sup.6]) Time-mean of age (*100) -0.014 (0.1320) -0.003 Time-mean of married 0.358 (0.100) *** 0.088 Time-mean of divorced 0.435 (0.124) *** 0.107 Time-mean of high school 0.357 (0.186) * 0.088 Time-mean of bachelor 0.392 (0.2900) 0.096 Time-mean of more than 0.758 (0.376) ** 0.186 bachelor Time-mean of urban 0.048 (0.0860) 0.012 Time-mean of limited -0.732 (0.098) *** -0.180 Dependent variable is a dummy if the individual has a positive log wage, 0 otherwise. Additional variables included in the models but not reported are three region dummies, four categorical variables for unemployment rate, and year dummies. Selection equations in columns 3 and 4 also include time-means for these variables. * Significant at 10% level. ** Significant at 5% level. *** Significant at 1% level. Table A2 Coefficient Estimates from Second-Stage Wage Equations OLS on the Pooled Sample Males Females Depressive symptoms -0.136 (0.018) *** -0.035 (0.013) *** Age -0.002 (0.003) -0.008 (0.003) ** Black -0.132 (0.017) -0.013 (0.016) Hispanic -0.060 (0.018) 0.007 (0.018) Married 0.157 (0.018) *** 0.015 (0.014) Divorced 0.069 (0.022) *** 0.083 (0.019) *** High school 0.170 (0.019) *** 0.120 (0.023) *** Bachelor 0.462 (0.029) *** 0.405 (0.030) *** More than bachelor 0.626 (0.044) *** 0.510 (0.043) *** Experience 0.035 (0.006) *** 0.038 (0.004) *** Experience squared (*100) -0.022 (0.021) -0.005 (0.015) Urban 0.078 (0.014) *** 0.085 (0.014) *** Limited -0.378 (0.065) *** -0.119 (0.028) *** Union coverage 0.201 (0.014) *** 0.123 (0.015) *** Self-employed -0.033 (0.032) -0.188 (0.041) *** Inverse mills 1.805 (0.458) *** 0.336 (0.160) ** Mean-Difference (using years 2004, 1994, 1992) Males Females Depressive symptoms -0.032 (0.017) * -0.014 (0.015) Age 0.001 (0.033) -0.028 (0.039) Black Hispanic Married 0.055 (0.025) ** -0.011 (0.025) Divorced 0.060 (0.034) * 0.026 (0.027) High school -0.085 (0.072) 0.087 (0.083) Bachelor 0.100 (0.093) 0.202 (0.101) ** More than bachelor 0.374 (0.125) *** 0.376 (0.114) *** Experience 0.116 (0.013) *** 0.070 (0.011) Experience squared (*100) -0.141 (0.023) *** -0.078 (0.020) *** Urban 0.026 (0.018) 0.019 (0.018) Limited -0.138 (0.049) *** -0.026 (0.029) Union coverage 0.122 (0.019) *** 0.097 (0.020) *** Self-employed -0.053 (0.040) -0.244 (0.052) *** Inverse mills 0.230 (0.242) -0.128 (0.082) Dependent variable is the logarithm of hourly wages. Models include occupation and industry dummies, categorical variables for local unemployment rates, and year dummies. Coefficient estimates and standard errors of experience squared are multiplied by 100 for presentation purposes. Robust standard errors are in parentheses. * Significant at 10% level. ** Significant at 5% level. *** Significant at 1% level.
I am grateful to Glenn Blomquist, Christopher Bollinger, John Garen, James Matron, Frank Scott, Aaron Yelowitz, James P. Ziliak, three anonymous referees, and the editor for helpful comments and suggestions. I also benefited from comments of seminar participants at the Southern Economic Association's Annual Meetings in 2006. All remaining errors are the author's responsibility.
Received December 2006; accepted October 2007.
Alexandre, Pierre K, Joseph Y. Fede, and Marsha Mullings. 2004. Gender differences in the labor market effects of serious mental illness. In The economics of gender and mental illness, edited by Dave E. Marcotte and Virginia Wilcox-G6k. Amsterdam: Elsevier Science, pp. 53-71.
Alexandre, Pierre K., and Michael T. French. 2001. Labor supply of poor residents in metropolitan Miami, Florida: The role of depression and the co-morbid effects of substance use. Journal of Mental Health Policy and Economics 4:161-73.
Baldwin, Marjorie L., and Steven C. Marcus. 2006. Labor market outcomes of persons with mental disorders. Paper presented at the Inaugural Conference of the American Society of Health Economics: Economics of Population Health. Madison, WI: 5 June 2006.
Barrick, Murray R., and Michael K. Mount. 1991. The big five personality dimensions and job performance: A rectaanalysis. Personnel Psychology 44:1-26.
Bartel, Ann, and Paul Taubman. 1979. Health and labor market success: The role of various diseases. Review of Economics and Statistics 61:1-8.
Bartel, Ann, and Paul Taubman. 1986. Some economic and demographic consequences of mental illness. Journal of' Labor Economics 4:243 56.
Berndt, Ernst R., Stan N. Finkelstein, Paul E. Greenberg, Robert H. Howland, Alison Keith, A. John Rush, James Russell, and Martin B. Keller. 1998. Workplace performance effects from chronic depression and its treatment. Journal of Health Economies 17:511-35.
Bono, Joyce E., and Timothy A. Judge. 2003. Core self-evaluations: A review of the trait and its role in job satisfaction and job performance. European Journal of Personality 17(Suppl.):S5-18.
Bound, John, David A. Jaeger, and Regina M. Baker. 1995. Problems with instrumental variables estimation when the correlation between the instruments and the endogenous explanatory variable is weak. Journal of the American Statistical Association 90:443-50.
Bray, Jeremy W. 2005. Alcohol use, human capital, and wages. Journal of Labor Economies 23:279-312.
Brown, G. W., and T. O. Harris. 1986. Establishing causal links: The Bedford College studies of depression. In Life events and psychiatric disorders." Controversial issues, edited by H. Katschnig. Cambridge, UK: Cambridge University Press, pp. 107-87.
Clark, Andrew E., Yannis Georgellis, and Peter Sanfey. 2001. Scarring: The psychological impact of past unemployment. Economica 68:221-41.
Clark, Lee A., Jeffrey Vittengl, Dolores Kraft, and Robin B. Jarrett. 2003. Separate personality traits from states to predict depression. Journal of Personality Disorders 17:152-72.
Clark, Lee A., David Watson, and Susan Mineka. 1994. Temperament, personality, and the mood and anxiety disorders. Journal of Abnormal Psychology 103:103-16.
Conley, J. J. 1984. The hierarchy of consistency: A review and model of longitudinal findings on adult individual differences in intelligence, personality and self-opinion. Personality and Individual Differences 5:11-25.
Cornwell, Christopher, and Peter Rupert. 1997. Unobservable individual effects, marriage and the earnings of young men. Economic Inquiry 35:285-94.
Costello, Elizabeth J. 1982. Locus of control and depression in students and psychiatric outpatients. Journal of Clinical Psychology 38:340-3.
Depression Guideline Panel. 1993. Depression in primary care." Volume 1. Detection and diagnosis. Clinical Practice Guideline, Number 5. Rockville, MD: U.S. Department of Health and Human Services, Public Health Service, Agency for Health Care Policy and Research. AHCPR Publication No. 93-0550.
Ettner, Susan L., Richard G. Frank, and Ronald C. Kessler. 1997. The impact of psychiatric disorders on labor market outcomes. Industrial and Labor Relations Review 51:64-81.
Frank, Richard, and Paul Gertler. 1991. An assessment of measurement error bias for estimating the effect of mental distress on income. Journal of Human Resources 26:154-64.
Goldsmith, Arthur H., Jonathan R. Veum, and William Darity, Jr. 1997. The impact of psychological and human capital on wages. Economic Inquiry 35:815-29.
Goodstein, Leonard D., and Richard I. Lanyon. 1999. Applications of personality assessment to the workplace: A review. Journal of Business and Psychology 13:291-322.
Greenberg, Paul E., Ronald C. Kessler, Howard G. Birnbaum, Stephanie A. Leong, Sarah W. Lowe, Patricia A. Berglund, and Patricia K. Corey-Lisle. 2003. The economic burden of depression in the United States: How did it change between 1990 and 2000? Journal of Clinical Psychiatry 62:1465-75.
Griliches, Zvi, and Jerry A. Hausman. 1986. Errors in variables in panel data. Journal of Econometrics 31:93-118.
Grzywacz, Joseph G., Joseph D. Hovey, Laura D. Seligman, Thomas A. Arcury, and Sara A. Quandt. 2006. Evaluating short-form versions of the CES-D for measuring depressive symptoms among immigrants from Mexico. Hispanic Journal of Behavioral Sciences 28:404-24.
Gustavsson, J. Petter, Robert M. Weinryb, Sara Goransson, Nancy L. Pedersen, and Marie Asberg. 1997. Stability and predictive ability of personality traits across 9 years. Personality and Individual Differences 22:783-91.
Hettema, Joop. 1995. Personality and depression: A multilevel perspective. European Journal of Personality 9:401-12.
Hogan, Robert, Joyce Hogan, and Brent W. Roberts. 1996. Personality measurement and employment decisions: Questions and answers. American Psychologist 51:469-77.
Hotz, V. Joseph, Susan W. McElroy, and Seth G. Sanders. 2005. Teenage childbearing and its life cycle consequences: Exploiting a natural experiment. Journal of Human Resources 40:683-715.
Intercontinental Marketing Services Health. 2006. Top-Line Industry Data. Accessed 25 October 2006. Available www. imshealth.com.
Jofre-Bonet, Mireia, Susan H. Busch, Tracy A. Falba, and Jody g. Sindelar. 2005. Poor mental health and smoking: Interactive impact on wages. Journal of Mental Health Policy and Economics 8:193-203.
Kendler, Kenneth S., Laura M. Karkowski, and Carol A. Prescott. 1999. Causal relationship between stressful life events and the onset of major depression. American Journal of Psychiatry 156:837-41.
Kessler, Ronald C., Catherine Barber, Howard G. Birnbaum, Richard G. Frank, Paul E. Greenber, Robert M. Rose, Gregory E. Simon, and Phillip Wang. 1999. Depression in the workplace: Effects on short-term disability. Health Affairs 18:163-71.
Kessler, Ronald C., Patricia Berglund, Olga Demler, Robert Jin, Doreen Koretz, Kathleen R. Merikangas, A. John Rush, Ellen E. Wakers, and Phillip S. Wang. 2003. The epidemiology of major depressive disorder: Results from the National Comorbidity Survey Replication (NCS-R). Journal of the American Medical Association 289:3095-105.
Kessler, Ronald C., and Richard G. Frank. 1997. The impact of psychiatric disorders on work loss days. Psychological Medicine 27:861-73.
Levine, Phillip B., Tara A. Gustafson, and Ann D. Velenchik. 1997. More bad news for smokers? The effects of cigarette smoking on wages. Industrial and Labor Relations Review 50:493-509.
Lim, Debbie, Kristy Sanderson, and Gavin Andrews. 2000. Lost productivity among full-time workers with mental disorders. Journal of Mental Health Policy and Economics 3:139-46.
Marcotte, Dave E., and Virginia Wilcox-Gok. 2001. Estimating the employment and earnings costs of mental illness: Recent developments in the United States. Social Science and Medicine 53:21-7.
Marcotte, Dave E., Virginia Wilcox-Gok, and D. Patrick Redmon. 1999. Prevalence and patterns of major depressive disorder in the United States labor force. Journal of Mental Health Policy and Economics 2(3): 123-31.
Marcotte, Dave E., Virginia Wilcox-Gok, and D. Patrick Redmon. 2000. The labor market effects of mental illness: The case of affective disorders. In The economics of disability, edited by D. Salkever. Greenwich, CT: JAI Press, pp. 181-210.
McCrae, Robert R., and Paul T. Costa, Jr. 1994. The stability of personality: Observations and evaluations. Current Directions in Psychological Science 3:173-5.
Melchior, L. A., G. J. Huba, V. B. Brown, and C. J. Reback. 1993. A short depression index for women. Educational and Psychological Measurement 53:1117-25.
Monroe, Scott M., and Anne D. Simons. 1991. Diathesis-stress theories in the context of life stress research: Implications for the depressive disorders. Psychological Bulletin 110:406-25.
Mundlak, Yair. 1978. On the pooling of time series and cross section data. Econometrica 46:69-85.
Paterniti, Sabrina, Isabelle Niedhammer, Thierry Lang, and Silla M. Consoli. 2002. Psychosocial factors at work, personality traits and depressive symptoms: Longitudinal results from the GAZEL Study. British Journal of Psychiatry 181:111-7.
Pearlin, Leonard I., Morton A. Lieberman, Elizabeth G. Menaghan, and Joseph T. Mullan. 1981. The stress process. Journal of Health and Social Behavior 22:337-56.
Radloff, Lenore S. 1977. The CES-D scale: A self-report depression scale for research in the general population. Applied Psychological Measurement 1:385-401.
Robins, Clive J. 1995. Personality-event interaction models of depression. European Journal of Personality 9:367-78.
Rosenberg, Morris. 1965. Society and the adolescent self-image. Princeton, NJ: Princeton University Press.
Ross, Catherine E., and John Mirowsky. 1989. Explaining the social patterns of depression: Control and problem solving--Or support and talking? Journal of Health and Social Behavior 30:206-19.
Rotter, Julian B. 1966. Generalized expectancies for internal versus external control of reinforcement. Psychological Monographs: General and Applied 80:1-28.
Rouch-Leroyer, Isabelle, Christophe Sourgen, Pascale Barberger-Gateau, Rebecca Fuhrer, and Jean-Francois Dartigues. 2000. Detection of depressive symptomatology in elderly people: A short version of the CES-D scale. Aging: Clinical and Experimental Research 12:228-33.
Santor, D. A. 1998. The Centre for Epidemiologic Studies depression scale: Working paper. Unpublished paper, Dalhousie University.
Stewart, Walter F., Judith A. Ricci, Elsbeth Chee, Steven R. Hahn, and David Morganstein. 2003. Cost of lost productive work time among U.S. workers with depression. Journal of the American Medical Association 289:3135-44.
Tett, Robert P., Douglas N. Jackson, and Mitchell Rothstein. 1991. Personality measures as predictors of job performance: A meta-analytic review. Personnel Psychology 44:703-42.
U.S. Department of Health and Human Services. 1999. Mental Health." A Report of the Surgeon General. Accessed 25 October 2006. Available http://www.surgeongeneral.gov/library/mentalhealth/home.html. Watson, David, and Lee A. Clark. 1995. Depression and the melancholic temperament. European Journal of Personality 9:351-66.
Wooldridge, Jeffrey M. 2002. Econometric analysis of cross section and panel data. Cambridge, MA: MIT Press.
World Health Organization. 2001. The world health report: 2001. Mental health: New understanding, new hope. Geneva, Switzerland: World Health Organization.
(1) Section 4 discusses the CES-D questionnaire and how the presence of depressive symptoms is determined. Description of the NLSY79 can also be found in section 4.
(2) Questions about depressive symptoms were asked in 1992, 1994 (the early period), and once in 1998, 2000, 2002, or in 2004 (the late period), depending on when the individual turned 40 years old.
(3) Ninety-five percent confidence intervals were applied but are not shown here in order to keep the figure easier to interpret and read. The 95% confidence intervals, however, reveal that those without symptoms have significantly larger wages from 1982, and those with depressive symptoms in both periods have significantly lower wages than the middle two groups from 1991. The confidence intervals always overlap for the two middle lines, though.
(4) Their OLS results for depression are positive and insignificant for both men and women.
(5) Studies covered in the paper mainly analyze personality dimensions that are important for vulnerability to depression, such as neuroticism and extraversion.
(6) Most respondents in the sample in the fixed-effects model are 30 years of age or older.
(7) For men, death of a child is not significant in the early years (1992 and 1994); however, it is highly significant in 2004. For women, death of a child is only significant in 1992. The presence of thyroid problems and the death variables in 2004 are jointly significant in the first stage for both men and women. For women, the inclusion of change in menstrual patterns, and especially happiness in marriage, makes the instruments jointly more significant; however, there is no substantial improvement in the predictive power in the first stages, and the partial [R.sup.2] is still low (around 0.04).
(8) Simultaneity could be more complicated, as one of the referees pointed out. For instance, wages could affect depression with a lag. To check whether this lagged effect is present in the sample, I also estimate simple linear models to predict depression with earlier log wage terms included in the model. When log wage from the previous year is included along with the current log wage, the coefficient of this latter is always stronger, both in magnitude and significance. When log wages from the two previous years are also included, only the current log wage remains a significant predictor of depressive symptoms.
(9) The CES-D was developed by the National Institute for Mental Health's Center for Epidemiologic Studies.
(10) Sixteen is the most frequently used cut-off, but alternative cut-off points have been also suggested (see Santor 1998).
(11) This Health Module is administered to respondents above the age of 40 beginning with the 1998 survey. Depending on when respondents turn 40 (or depending on availability for interview), these Health Module questions were asked in 1998, 2000, 2002, and 2004 (each respondent answered once only). For the sake of simplifying notation, I will refer to the results gained from using information provided in the 'Health Module 40 & Over' as 2004, the last year of data availability. These results, however, overlap four waves. These waves are controlled for in all of the models by the inclusion of year dummies.
(12) The seven questions asked in each of the years 1992, 1994, and 2004 are as follows: (1) I did not feel like eating; my appetite was poor; (2) I had trouble keeping my mind on what 1 was doing; (3) I felt depressed; (4) I felt that everything I did was an effort; (5) My sleep was restless; (6) I felt sad; and (7) I could not get "going." There are four response categories, depending on how often respondents had these emotions. The categories are 0 = "Rarely/None of the Time/1 Day"; 1 = "Some/A Little of the Time/1-2 Days"; 2 = "Occasionally/Moderate Amount of the Time/3-4 Days"; and 3 = "Most/All of the Time/5-7 Days."
(13) Distribution of the score throughout time is displayed in the Appendix (Figure A1).
(14) There are various reasons why I use this threshold. As I mentioned before, the full questionnaire was available in 1992. Therefore, using information in 1992, I regressed the full CES-D score on the score that is obtained by using only the seven items that are available in multiple years. The estimated coefficient is 2.199, and since the most frequently used cut-off value is 16 for the full CES-D scale, this translates into a cut-off value of 7.27 (= 16/2.199) for the short form. Since only whole numbers can be used, seven seemed to be a reasonable cut-off value. The two measures agree over 90% of time in how they classify the sample--93% of the non-symptom group (as defined by the full scale) is in the non-symptom group, as defined by the short-scale, and 80% of those for whom the presence of depression is predicted by the full scale are also predicted by the short scale to have depressive symptoms. The Chi-square test of independence strongly rejects the null hypothesis that the two categorical variables are independent. Also, using the full-item CES-D questionnaire in 1992 indicates that about 21% of the sample show signs of depression. Using the short form with a cut-off of seven indicates that about 22% show signs of depression, while using a cut-off of six or a cut-off of eight shows signs of depression for about 29% and 17% of the sample, respectively. Furthermore, with a cutoff value of seven, the prevalence rates of depressive symptoms in the sample in different years are close to the combined prevalence rates of unipolar major depression, dysthymia, and "depression not otherwise specified" in the United States (Depression Guideline Panel 1993).
(15) Furthermore, people with real hourly wages less than $1 or greater than $1000 are also dropped. I also estimated models including those that had no meaningful CES-D, limitation, and non-labor income values by replacing missing values with zeros (or the average in the case of non-labor income) and including dummy variables controlling for respondents with missing values. Results are not significantly different from those presented below.
(16) Results are robust to the inclusion of those in the selection equation who have interpretable information in only one year of the analysis. (For these individuals, time-means of the explanatory variables in the selection equation are the same as the explanatory variables themselves.)
(17) There is no good physical health measure that is consistently available in the NLSY79. In the different model specifications, I use a dummy variable that takes on 1 if health limits the amount or kind of work of which the respondent is capable. Since the question is not specific to the source of this limitation, this variable could be related to mental health as well.
(18) The inverse Mills ratio is significant in all specifications except for women in 1992.
(19) Depressive symptoms is a dummy variable here; therefore, the adequate way to calculate the impact on wages in percentage is [MATHEMATICAL EXPRESSION NOT REPRODUCIBLE IN ASCII].
(20) In 1992, I can check the reliability of the symptom dummy based on the short CES-D, since the full scale is also available in this year. A dummy created with the usual cut-off value of 16 for the full version of CES-D has a coefficient estimate of -0.117 in the sample of men and -0.028 in the sample of women. Neither of these estimates is significantly different from those presented in Table 2 for the year 1992.
(21) For women, the parabola turns around a score of 13 and 14. For men, there is no turn in the relevant range of scores.
(22) Coefficient estimates of the 2004 symptoms in 1992 wage equations are -0.135 for men and -0.029 for women. In 1994 wage equations the estimates are 0.115 for men and -0.062 for women. The coefficient estimate for women in 1992 is insignificant.
(23) A higher score on the Rotter scale is associated with a more external locus of control. As for the Rosenberg scale, the higher the score, the higher the self-esteem.
(24) To preserve space, the table only shows the 2004-1992 differences. Differences using other possible years are similar in that change in logarithm of wages for those who show symptoms in the earlier year but not in the later period is not significantly different from those who show symptoms in the later period but not in the earlier years. Means of changes in the limitation dummy, however, are always significantly different for these groups for both males and females.
(25) The number of respondents with symptom changes in the female sample is 525 in the long-difference model, 540 in the short-difference model, and 849 in the mean-difference model.
(26) Average CES-D scores of the working group (non-working group) are between 2.497-3.485 (4.420-5.575) for men and 3.542-4.396 (5.338-5.474) range for women.
(27) Coefficient estimates of the symptom dummy in the pooled samples can be found in the first row of columns 1 and 2 in Table A1 in the Appendix.
(28) Similarly, first-stage selection equations using the scores reveal that current score and its squared term are never significant, but the time-mean of the score and the time-mean of the squared term significantly affect the probability of work in three out of the six specifications.
Attila Cseh, Department of Marketing and Economics, Harley Langdale, Jr. College of Business Administration, Valdosta State University, 1500 N. Patterson St., Valdosta, GA 31698, USA; E-mail firstname.lastname@example.org.
Table 1. Descriptive Statistics Men 1992 1994 2004 (a) Depressive symptoms (b) 0.174 (0.379) 0.137 (0.344) 0.132 (0.339) CES-D score 3.627 (3.637) 3.000 (3.469) 2.758 (3.815) Employed 0.875 (0.330) 0.882 (0.323) 0.915 (0.278) Logarithm of hourly (real) wage (c) 2.609 (0.551) 2.681 (0.561) 2.829 (0.640) Age 31.293 (2.204) 33.308 (2.202) 40.954 (0.876) Black 0.274 (0.446) 0.267 (0.442) 0.279 (0.448) Hispanic 0.192 (0.394) 0.192 (0.394) 0.190 (0.392) Married 0.564 (0.496) 0.583 (0.493) 0.602 (0.490) Divorced 0.085 (0.279) 0.109 (0.311) 0.156 (0.363) Children between 0 and 5 0.455 (0.716) 0.410 (0.675) 0.170 (0.455) Children between 6 and 18 0.548 (0.929) 0.678 (1.019) 1.072 (1.232) Less than high school 0.154 (0.361) 0.139 (0.346) 0.122 (0.327) High school 0.648 (0.478) 0.652 (0.476) 0.669 (0.471) Bachelor 0.151 (0.358) 0.151 (0.358) 0.141 (0.348) More than bachelor 0.047 (0.211) 0.058 (0.234) 0.068 (0.251) Experience 10.080 (3.522) 11.883 (3.824) 18.167 (5.287) Urban 0.769 (0.421) 0.774 (0.418) 0.754 (0.477) Limited 0.061 (0.239) 0.066 (0.248) 0.099 (0.299) N 3660 3427 3143 Women 1992 1994 2004 (a) Depressive symptoms (b) 0.260 (0.439) 0.255 (0.436) 0.206 (0.405) CES-D score 4.633 (4.239) 4.422 (4.381) 3.826 (4.466) Employed 0.780 (0.414) 0.782 (0.413) 0.845 (0.362) Logarithm of hourly (real) wage (c) 2.404 (0.548) 2.436 (0.591) 2.553 (0.610) Age 31.406 (2.197) 33.384 (2.205) 40.915 (0.827) Black 0.279 (0.449) 0.288 (0.453) 0.293 (0.455) Hispanic 0.195 (0.396) 0.193 (0.395) 0.193 (0.395) Married 0.564 (0.496) 0.571 (0.495) 0.571 (0.495) Divorced 0.125 (0.330) 0.141 (0.348) 0.192 (0.394) Children between 0 and 5 0.491 (0.701) 0.432 (0.672) 0.131 (0.389) Children between 6 and 18 1.016 (1.116) 1.191 (1.169) 1.468 (1.203) Less than high school 0.126 (0.331) 0.118 (0.322) 0.096 (0.295) High school 0.676 (0.468) 0.679 (0.467) 0.688 (0.464) Bachelor 0.158 (0.365) 0.158 (0.365) 0.153 (0.360) More than bachelor 0.040 (0.196) 0.045 (0.208) 0.063 (0.243) Experience 8.627 (4.037) 10.035 (4.533) 15.425 (6.271) Urban 0.804 (0.397) 0.797 (0.402) 0.740 (0.472) Limited 0.089 (0.284) 0.091 (0.287) 0.128 (0.334) N 3768 3556 3285 (a) 'Health Module 40+' questions were asked in 1998, 2000, 2002, and 2004. Each respondent was asked once depending on when he/she turned 40. In tables, I refer to the last year during which the Health Module 40+ was made publicly available, 2004. (b) I consider the presence of symptoms of depression if the sum of the CES-D score is greater than 7 (the maximum is 21). (c) Means of logarithm of hourly wages are based on the subsample that had non-missing values. Table 2. OLS Estimates in Wage Equations for Males and Females 1992 1994 Men A: Depressive symptoms - 0.146 (0.029) *** -0.124 (0.029) *** B: Score (*10) -0.071 (0.057) -0.198 (0.060) *** Score squared (*1000) -0.756 (0.512) 0.516 (0.486) N 3204 3022 Women A: Depressive symptoms -0.020 (0.019) -0.057 (0.021) *** B: Score (*10) -0.104 (0.051) ** -0.128 (0.054) ** Score squared (*1000) 0.416 (0.339) 0.489 (0.347) N 2940 2781 2004 Pooled (a) Men A: Depressive symptoms -0.157 (0.033) *** -0.136 (0.018) *** B: Score (*10) -0.294 (0.064) *** -0.200 (0.036) *** Score squared (*1000) 0.100 (0.442) 0.444 (0.275) N 2877 9103 Women A: Depressive symptoms -0.039 (0.025) -0.035 (0.013) *** B: Score (*10) -0.051 (0.057) -0.090 (0.033) *** Score squared (*1000) 0.125 (0.391) 0.342 (0.210) N 2777 8498 Dependent variable is the logarithm of hourly wages. Models include race/ethnic dummies (2), age, marital status dummies (2), dummies for geographic regions (3), experience, experience squared, limitation, urban residence, four categories for education, union status, a dummy for self-employment, occupation dummies (7), industry dummies (7), categorical variables for local unemployment rates, and year dummies in 2004 and in the column showing the pooled estimates. Models correct for selection into labor market. Model of selection includes non-labor income, number of small children, and number of older children. Number of observations in the selection equations for males (females) are 3660 (3768) in 1992.3427 (3556) in 1994, 3143 (3285) in 2004, and 10,230 (10,609) in the pooled specification. Coefficient estimates and standard errors of score squared are multiplied by 1000 for presentation purposes. Standard errors are in parentheses. (a) For more coefficient estimates of the selection equations and the second stage for the pooled sample, see the Appendix. * Significant at 10% level. ** Significant at 5% level. *** Significant at 1% level. Table 3. Coefficient Estimates in Wage Equations Using Average CES-D Scores and the Difference from the Average 1992 Men Average score -0.030 (0.005) *** Difference from average score 0.001 (0.004) N 2607 Women Average score -0.008 (0.003) ** Difference from average score 0.001 (0.003) N 2503 1994 Men Average score -0.023 (0.004) *** Difference from average score -0.001 (0.004) N 2482 Women Average score -0.008 (0.003) ** Difference from average score 0.000 (0.003) N 2385 2004 Men Average score -0.029 (0.004) *** Difference from average score 0.001 (0.004) N 2701 Women Average score -0.008 (0.003) *** Difference from average score -0.001 (0.003) N 2660 See notes to Table 2. ** Significant at 5% level. *** Significant at 10% level. Table 4. Comparing Personality Characteristics for Men and Women in Different Emotional States Depressed in the Early Never Period, Not Depressed Depressed (1) in the Late Period (2) Men Rotter 8.689 (2.041) 9.116 (2.087) Rosenberg-80 33.060 (3.946) 31.778 (3.979) Rosenberg-87 34.324 (3.953) 32.665 (4.101) Women Rotter 8.698 (2.032) 9.186 (2.144) Rosenberg-80 32.846 (4.028) 31.675 (3.958) Rosenberg-87 33.972 (3.946) 32.657 (4.115) Depressed in the Late Depressed in Both Period, Not Depressed the Early Period and in in the Early Period (3) the Late Period (4) Men Rotter 9.250 (2.044) 9.595 (2.134) Rosenberg-80 31.583 (4.123) 30.788 (4.138) Rosenberg-87 32.803 (4.060) 31.140 (4.262) Women Rotter 9.034 (2.204) 9.537 (2.136) Rosenberg-80 31.953 (3.950) 30.540 (3.968) Rosenberg-87 32.784 (4.319) 31.483 (4.072) Means in columns I and 4 are significantly different from those in all other columns (at least at the 10% level) for both men and women. Means in columns 2 and 3 are never significantly different from each other. (Standard deviations are in parentheses.) Table 5. Means and Standard Deviations of Changes in Certain Variables between 1992 and 2004 Symptom in 2004, No Symptom No Symptoms in 1992 (a) (b) Males [DELTA]CES-D score -0.754 (2.260) 7.129 (4.210) *** [DELTA]lnwage 0.242 (0.553) 0.141 (0.514) ** [DELTA]married 0.043 (0.493) -0.026 (0.522) [DELTA]divorced 0.056 (0.362) 0.084 (0.360) [DELTA]limited 0.014 (0.237) 0.181 (0.476) *** [DELTA]experience 8.811 (2.292) 8.527 (2.460) [DELTA]union 0.011 (0.401) 0.013 (0.395) N 1709 155 Females [DELTA]CES-D score -0.807 (2.440) 7.452 (4.088) *** [DELTA]lnwage 0.212 (0.569) 0.214 (0.516) [DELTA]married 0.025 (0.475) -0.025 (0.572) [DELTA]divorced 0.048 (0.394) 0.090 (0.440) [DELTA]limited 0.006 (0.237) 0.111 (0.424) *** [DELTA]experience 8.217 (2.386) 7.651 (2.606) *** [DELTA]union 0.029 (0.410) -0.030 (0.375) ** N 1313 199 No Symptom in 2004, Symptom in Both Symptom in 1992 2004 and 1992 (c) (d) Males [DELTA]CES-D score -7.126 (3.089) *** 0.462 (4.593) ** [DELTA]lnwage 0.226 (0.505) 0.168 (0.586) [DELTA]married 0.076 (0.527) 0.031 (0.467) [DELTA]divorced 0.084 (0.382) 0.092 (0.423) [DELTA]limited -0.046 (0.287) *** 0.015 (0.414) [DELTA]experience 8.567 (2.448) 7.664 (2.507) *** [DELTA]union -0.015 (0.401) -0.062 (0.527) N 262 65 Females [DELTA]CES-D score -7.224 (3.745) *** 0.383 (4.297) *** [DELTA]lnwage 0.241 (0.552) 0.135 (0.630) [DELTA]married 0.012 (0.555) -0.035 (0.513) [DELTA]divorced 0.080 (0.464) 0.135 (0.537) [DELTA]limited -0.031 (0.349) * 0.078 (0.464) * [DELTA]experience 7.893 (2.565) ** 7.748 (2.752) [DELTA]union 0.025 (0.429) 0.007 (0.485) N 326 141 p-Value of Testing the Difference of the Means in Columns b and c (e) Males [DELTA]CES-D score <0.0000 [DELTA]lnwage [DELTA]married 0.0551 [DELTA]divorced [DELTA]limited <0.0000 [DELTA]experience [DELTA]union N Females [DELTA]CES-D score <0.0000 [DELTA]lnwage [DELTA]married [DELTA]divorced [DELTA]limited 0.0001 [DELTA]experience [DELTA]union N Means in columns b, c, and d are tested against means in column a, the group without any symptoms. Column e shows the p-values of the tests of differences in means of columns b and c only if differences are significant at any of the usual levels of significance. * Significant at 10% level. ** Significant at 5% level. *** Significant at 1% level. Table 6. Fixed Effects Results for Men and Women Long Difference (2004-1992) (1) (2) Males A Depressive symptoms -0.030 (0.025) -0.030 (0.025) No. of changes in 417 depressive symptoms in the working sample B Score (*10) -0.040 (0.060) -0.067 (0.063) Score squared (*1000) -0.122 (0.458) 0.101 (0.474) Females A Depressive symptoms -0.015 (0.023) -0.014 (0.024) No. of changes in 525 depressive symptoms in the working sample B Score (*10) 0.030 (0.058) 0.017 (0.060) Score squared (*1000) -0.258 (0.358) -0.227 (0.365) All control variables Yes No included Short Difference (1994-1992) (3) (4) Males A Depressive symptoms -0.027 (0.022) -0.044 (0.022) ** No. of changes in 451 depressive symptoms in the working sample B Score (*10) 0.009 (0.064) -0.063 (0.065) Score squared (*1000) -0.578 (0.548) -0.240 (0.607) Females A Depressive symptoms -0.032 (0.019) * -0.025 (0.020) No. of changes in 540 depressive symptoms in the working sample B Score (*10) -0.082 (0.054) -0.092 (0.053) * Score squared (*1000) 0.501 (0.347) 0.555 (0.339) All control variables Yes No included Mean Difference (using years 2004, 1994, 1992) (a) (5) (6) Males A Depressive symptoms -0.032 (0.017) * -0.035 (0.017) ** No. of changes in 703 depressive symptoms in the working sample B Score (*10) -0.005 (0.043) -0.013 (0.043) Score squared (*1000) -0.345 (0.340) -0.369 (0.361) Females A Depressive symptoms -0.014 (0.015) -0.020 (0.015) No. of changes in 849 depressive symptoms in the working sample B Score (*10) 0.003 (0.038) -0.006 (0.039) Score squared (*1000) -0.047 (0.236) -0.048 (0.237) All control variables Yes No included Coefficient estimates in columns 1, 3, and 5 were obtained from models including a full set of control variables. Models in columns 2, 4, and 6 exclude all control variables, with the exception of the main variable of interest and year dummies. Estimates of score and score squared are jointly significant for males in column 3 (at the 10% level of significance), in column 4 (at the 1% level of significance), in column 5 (at the 10% level of significance), and in column 6 (at the 5% level of significance). (a) See the Appendix for more coefficient estimates of the mean-difference model estimating the effect of depressive symptoms. * Significant at 10% level. ** Significant at 5% level.
|Printer friendly Cite/link Email Feedback|
|Comment:||The effects of depressive symptoms on earnings.|
|Publication:||Southern Economic Journal|
|Date:||Oct 1, 2008|
|Previous Article:||Comparing small-group and individual behavior in lottery-choice experiments.|
|Next Article:||Multiple-output child health production functions: the impact of time-varying and time-invariant inputs.|