- Email: [email protected]
Breast cancer survival, work, and earnings
Journal of Health Economics 21 (2002) 757–779
Breast cancer survival, work, and earnings Cathy J. Bradley a,∗ , Heather L. Bednarek b , David Neumark c a
Department of Medicine, College of Human Medicine, B212 Clinical Center, Michigan State University, East Lansing, MI 48824, USA b Department of Economics, St. Louis University, Saint Louis, MO 63108, USA c Department of Economics, Michigan State University, East Lansing, MI 48824, USA Received 1 May 2001; accepted 1 February 2002
Abstract Relying on data from the Health and Retirement Study (HRS) linked to longitudinal social security earnings data, we examine differences between breast cancer survivors and a non-cancer control group in employment, hours worked, wages, and earnings. Overall, breast cancer has a negative impact on employment. However, among survivors who work, hours of work, wages, and earnings are higher compared to women in the control group. We explore possible biases underlying these estimates, focusing on selection, but cannot rule out a causal interpretation. Our research points to heterogeneous labor market responses to breast cancer, and shows that breast cancer does not appear to be debilitating for women who remain in the work force. © 2002 Elsevier Science B.V. All rights reserved. JEL classification: I1; E61; J21 Keywords: Breast cancer; Employment; Labor market effects; Hours worked; Earnings
1. Introduction In recent years, improved detection methods for breast cancer have led to the treatment and survival of a younger population of women more likely to be at working ages, making an inquiry into the impact of breast cancer on labor market outcomes particularly relevant. Between 1983 and 1993, in situ breast cancer rates increased from 2.3 to 6.2 per 100,000 among women under age 50, largely reflecting an increase in the use of mammography (American Cancer Society, 1999). Treatment has improved as well, leading to the largest short-term decline in over 40 years in breast cancer mortality (American Cancer Society, ∗ Corresponding author. Tel.: +1-517-432-3405; fax: +1-517-432-9471. E-mail address: [email protected] (C.J. Bradley).
0167-6296/02/$ – see front matter © 2002 Elsevier Science B.V. All rights reserved. PII: S 0 1 6 7 - 6 2 9 6 ( 0 2 ) 0 0 0 5 9 - 0
758
C.J. Bradley et al. / Journal of Health Economics 21 (2002) 757–779
2001). Breast cancer research has focused on detection and treatment and, to a lesser extent, survivors’ quality-of-life. However, now that 5-year survival is expected for most women diagnosed in the early stages of breast cancer, attention should also be given to economic measures of the consequences of surviving breast cancer, as part of a broader effort to understand the quality-of-life implications of the disease, and how labor market agents react to cancer survivorship. We focus on breast cancer in this paper because we believe that it is in some important respects unique and worthy of study as a single disease. Breast cancer mostly affects women, and screening for breast cancer (i.e. mammography) is routinely applied to working age women, yet the long-term benefits of screening and treatment of early stage cancers are unknown (Olsen and Gotzsche, 2001), making an inquiry into breast cancer’s effect on productivity particularly relevant. Our view is that other cancers do not share breast cancer’s characteristics in terms of its screening, treatment, and prognosis, and that few other diseases have the same emotional impact as “cancer.” In this paper, we use the first wave of the Health and Retirement Study (HRS)—in some cases linked to longitudinal social security earnings data—in an attempt to understand how breast cancer influences labor market decisions and outcomes. As cancer screening is more routinely applied to a working age population, more cancers are likely to be detected in early stages that may have otherwise gone undetected until a later time. Therefore, this research is particularly relevant as it both fills a gap in the literature regarding labor market consequences of cancer and provides information on how early detection of disease may affect labor market outcomes.
2. Illness and labor market outcomes Intuitively, poor health would seem to have a negative impact on labor supply and productivity. A more formal way of thinking about this impact is that under the assumption of utility maximization, the supply of work hours H (the difference between a time endowment T and the demand for leisure hours L) is determined by tastes, prices, and endowments (e.g. of wealth). Poor health can affect labor supply by diminishing tastes for work and thereby raising the marginal value of leisure time, reducing productivity, and, put simply, by stealing time away from work for health maintenance (Grossman, 1972). Formally, this could be captured by incorporating a health production function into a labor supply model. Thinking beyond this basic model, the potential for a change in health insurance status may be an important consideration in the decision to exit the labor force, given that breast cancer treatment is expensive and the potential for future medical expenses is great. Under these circumstances, an individual’s tolerance for financial and personal risk becomes important. The link between health insurance and employment, particularly in the United States, may create unusual incentives regarding the decision to work (Currie and Madrian, 1999). For people with medical conditions who need specialized (and often expensive) care, obtaining coverage is uncertain (Friedland, 1996). Thus, some people may be trapped in a job for fear of losing insurance and others may have disincentives for leaving public programs and seeking employment since having a job can mean losing coverage (Adams, 2001; Madrian, 1994). Therefore, the counterintuitive notion that in the face of illness labor
C.J. Bradley et al. / Journal of Health Economics 21 (2002) 757–779
759
supply may actually increase or remain unchanged, through the influence of health insurance and its accompanying incentives, is plausible. This tension between health conditions and their negative influence on labor supply, and the demand for health insurance and its potentially positive influence on labor supply, is reflected in the existing empirical research. In empirical studies, the relationship between poor health or disability, health insurance, and labor supply appears unresolved (for an overview see Currie and Madrian, 1999). Psychiatric disorders, for example, have been shown to significantly reduce employment, work hours, and income (Ettner et al., 1997; Benham and Benham, 1982). Severity of illness also plays an important role in the decision to reduce participation in the labor force (Famulari, 1992). Conversely, there is some evidence consistent with the demand for health insurance increasing or maintaining labor supply in response to illness. An ill person— particularly someone with a chronic condition requiring specialized care—is motivated to remain at work to maintain health insurance coverage. Empirical studies (e.g. Gruber and Madrian, 1994; Madrian, 1994; Monheit and Cooper, 1994; Kapur, 1998) supporting this position largely address the potential for job lock that may decrease workers’ mobility, rather than directly addressing workers’ willingness to exit the labor force or curtail hours. Cancer-specific empirical studies have pointed to some key questions in understanding the direction of influence cancer is likely to have on labor market decisions and outcomes, but they are far from decisive, and do not focus in a very detailed manner on the empirical estimation of the effects of cancer on these decisions and outcomes. In general, studies (e.g. Quigley, 1989; Fow, 1996; Kornblith, 1998; Razavi et al., 1993; Carter, 1994; Berkman and Sampson, 1993) have focused on survivors’ subjective impressions of the impact of cancer on their lives. These studies suggest negative factors that can reduce employment, including physical disability (e.g. limitations in upper body strength; Fow, 1996), memory loss (Schagen et al., 1999), lack of control over schedules, need for transportation, and type of work performed (Greenwald et al., 1989; Satariano and DeLorenze, 1996), and in some cases discrimination on the part of employers (Carter, 1994; Berry, 1993). Other studies of employment following breast cancer suggest an alternative view. For example, in a preliminary study, we found that breast cancer survivors work more hours per week than a non-cancer control group (Bradley et al., in press). In a study of patients 2 and 3 years after their primary treatment, Ganz et al. (1996) found that breast cancer survivors function at a high level. Sixty-five percent of survivors (n = 139) were either working for pay or volunteering their services. The mean numbers of hours worked were 34.4 and 33.2 hours per week among women who were 2 and 3 years post-treatment, respectively. This study concluded that women continue to work and perform their usual roles after treatment for breast cancer. Satariano and DeLorenze (1996) had similar findings. In a study of 300 women working at the time of their breast cancer diagnosis, 71% returned to work within 3 months after diagnosis. The implication of these studies is that the overall impact of breast cancer on employment status is minor. However, these studies are not based on national samples, nor do they include a control group. Our analysis contributes to research on this topic along both of these dimensions. In addition, unlike much of the previous research, we control for a number of correlates that can influence the probability of employment, decisions regarding hours of work, and earnings, including individual characteristics (in some analyses unobserved characteristics) and information on health insurance coverage.
760
C.J. Bradley et al. / Journal of Health Economics 21 (2002) 757–779
A fundamental problem presented by our data set is that we analyze breast cancer survivors and a control group. This raises the possibility that differences we find between these two groups are attributable to non-random (conditional on the observables) differences between survivors and non-survivors—based on “selection” into the sample of survivors— rather than to actual effects of breast cancer. We try to address the issue as best as possible based on the observable information we have. In addition, though, when we look at earnings—perhaps the best summary measure of labor market outcomes—we are able to exploit a unique feature of our data set that allows us to link to longitudinal administrative data on earnings of breast cancer survivors and the control group. With these data, we use the pre-diagnosis observations on the survivors, and examine differences relative to earlier observations on the control group, to control for possible selection bias. 3. Data Our main data set is drawn from the first wave (1992) of the HRS. The HRS is designed to answer research questions regarding the relationship between health, income, wealth, job decisions, and retirement. The cohort is aged 51–61.1 In most cases, if there are two members of the household (e.g. husband and wife) each is interviewed regarding his/her own employment history, retirement, health, and demographic characteristics. The more financially knowledgeable respondent in the household is interviewed regarding housing, net worth, income, and health insurance status for the members of the household included in the survey.2 Only one member of the household must be in the age range of the sample frame; therefore spouses may be interviewed even if they are out of the specified age range, and thus our sample departs from national representativeness for these respondents. For our non-cancer control group, we selected women who answered “no” to the question, “Has a doctor ever told you that you have cancer or a malignant tumor of any kind?” Given that a cancer diagnosis is a life-altering event, recall bias is not likely to be a problem in the data. Those women that said “yes” to the above question and subsequently indicated that their diagnosis was breast cancer constitute the treatment group. We also consider information available in the HRS on the number of years since breast cancer diagnosis, which is likely to be associated with morbidity and the severity of any side effects from treatment, and therefore to have potential labor market consequences. Women in our sample were diagnosed an average of 7 years (standard deviation (S.D.) = 6, range = 1–32) prior to the interview. The HRS contains 6708 women on whom we have information regarding labor force participation and hours worked. We excluded women who were insured by Medicaid (n = 319) because the program has been shown to constrain labor supply, particularly for families with high expected medical expenditures (Moffitt and Wolfe, 1992). The non-Medicaid restriction coupled with missing data3 led to a final sample size of 5974 women, 156 of whom reported having breast cancer. In addition, to carry out the longitudinal analysis of 1
There are follow-up interviews at 2-year increments, which are not used in this study. If a person was single, he or she was designated the primary respondent. In a married pair or partnership, the person most knowledgeable of household financial matters was designated the primary respondent. 3 Exclusions and missing data were distributed as follows: Medicaid (319), age (120), type of cancer (282), race (13). 2
C.J. Bradley et al. / Journal of Health Economics 21 (2002) 757–779
761
earnings discussed earlier we obtained the restricted social security annual earnings file (covering years 1951–1991) for HRS respondents. Social security earnings were reported for 4527 women; 2936 (68 breast cancer survivors) of them worked in 1991.
4. Empirical approach 4.1. Specifications The outcomes of interest are employment status, usual hours of weekly work, wages, and earnings. Each of these can be expressed as a function of the incidence of breast cancer (BCA), the duration of survival to date (S), exogenous variables (X), health insurance (HI), and random or unobserved influences (ε). Generically, we write the employment equation as: Ei∗ = f (BCAi , Si , Xi , HIi , εi )
(1)
Ei∗
where represents a latent variable for the propensity for employment. We estimate the employment equation by assuming a linear additive form for Eq. (1),4 Ei∗ = αE + βE BCAi + Xi γE + δE HIi + εi
(2)
We then define employment status as a binary variable (Ei ) that equals one if the respondent reports positive hours worked and zero otherwise, and estimate Eq. (2) as a probit model. For ease of interpretation, the probit estimates are translated into derivatives of the probability of working with respect to the independent variables.5 Health insurance may affect labor supply as well as mediate the effects of breast cancer. It is available in large part, for the women in our sample, if they are employed or married with an employed spouse. It is highly likely that the presence of employer-based health insurance is endogenous to a model estimating employment, and so we approach the estimation of the effect of health insurance, for married women, using information on whether her spouse has health insurance through his employer (Buchmueller and Valletta, 1999; Olson, 1998). Consequently, when we estimate Eq. (2) including information on spousal health insurance, we use a sample of married women only. In our estimation of the effects of breast cancer on hours (H), wages (W), and earnings (Y), we assume that the same variables that affect employment also affect these dependent variables. We focus on the estimates of the conditional functions for: Hi = αH + βH BCAi + Xi γH + ηi ,
if Ei = 1,
Wi = αW + βW BCAi + Xi γW + µi , Yi = αY + βY BCAi + Xi γY + νi , 4
if Ei = 1,
if Ei = 1,
(3) (4) (5)
Note that we have omitted Si for now; later, we introduce information on length of survivorship. The “derivatives” are computed as the difference in probabilities as the dummy variable takes on the values zero and one, with the other variables at the sample means. The standard errors reported are those making the t-statistics for these derivatives the same as those on the original coefficients, allowing the reader to easily assess the statistical significance of the raw coefficients. 5
762
C.J. Bradley et al. / Journal of Health Economics 21 (2002) 757–779
recovering estimates of the linear projections of Hi , Wi , or Yi on the independent variables for those working.6,7,8,9 Hours are defined as average hours worked (H) per week while earnings (Y) are based on the respondent’s reported annual earnings in 1991 (and in some cases social security reported annual earnings). Wages (W) were reported on an hourly basis by many HRS respondents (n = 1956). For respondents who reported their earnings in weekly, bi-weekly, monthly, or annual increments, we computed wages by dividing their reported earnings by their average hours worked during the specified time frame. For example, if the respondent worked an average of 40 hours per week and earnings were reported bi-weekly, we divided bi-weekly earnings by 80.10 4.2. Econometric issues In order for a breast cancer survivor to appear in our data set, she must have survived breast cancer for some time. In the HRS sample, 76% of the patients had survived 3 or more years. It has been argued that labor market decisions and health should be jointly considered (Grossman, 1972; Olson, 1998). This is a potentially futile challenge given our data, since we know very little about respondents’ risk factors for breast cancer (e.g. family history, nulliparity, estrogen use, high fat diets, and environmental risks). Thus, treating the incidence of breast cancer as essentially random, other than a few known exogenous predictors out of individuals’ control such as age and race,11 seems a reasonable approximation to reality. Likewise, survival is threatened by cancer recurrences and metastases-events that may also be related to unobservable variables such as breast cancer stage at diagnosis, histology, and other comorbidities. A causal interpretation of the estimates of the equations described in the previous section requires that survival be unrelated to the unobservables.12 To fully address this issue, we would require longitudinal data on women before and after the onset of breast cancer (along with clinical information on their cancer). In the absence of such data and under weaker assumptions (that the unobserved residuals are fixed over time for individuals), we can recover causal estimates by comparing changes in behavior for 6
We assume a linear functional form for hours worked and a log-linear functional form for wages and earnings. We estimated Eqs. (3) and (5) for married women, controlling for whether her spouse had health insurance through an employer. However, the results did not qualitatively change. We do not estimate Eq. (4) in this manner, as spouse’s health insurance is unlikely to independently affect productivity. 8 Such regressions generally provide biased estimates of the slopes of the unconditional mean function (unless the conditional expectations of the error terms are uncorrelated with the regressors), but still are informative about how the exogenous variables are related to hours of work (wages and earnings) for those at work. The selection–correction model (Heckman, 1979) is inappropriate since there are no natural candidates for variables that could be excluded a priori from the second-stage equation, and identification would rely on functional form restrictions (e.g. linearity) or distributional assumptions (see Olsen, 1980). 9 In all cases we also estimated Tobit models, which are consistent if the normality assumption holds but not otherwise; throughout the empirical analysis, the results were not sensitive to the choice of estimator. Results are available upon request. 10 Extreme outliers in reported or computed wages were deleted, with no substantive impact on the results. 11 Caucasian women are slightly more likely to develop but also to survive breast cancer compared to other women. 12 Similarly, consistent estimation of the conditional mean specifications including S (the number of years of survival thus far) requires that this survival duration also be unrelated to the unobservables. 7
C.J. Bradley et al. / Journal of Health Economics 21 (2002) 757–779
763
cancer patients (pre- and post-diagnosis/survival) to changes in behavior for the non-cancer control group over a similar period. Such data are by and large absent in the HRS.13 However, one exception is the linked social security annual earnings records available for each of the up to 40 years prior to the interview, which can be obtained from the administrators of the HRS through special arrangements designed to protect respondents’ confidentiality. Therefore, we first address the issue of causality by exploiting the information in the social security data on earnings in the years prior to diagnosis, to conduct longitudinal analyses. This allows us to examine earnings pre- and post-diagnosis for breast cancer survivors and at comparable times for the control group, to see if post-diagnosis differentials between breast cancer survivors and the control group also existed prior to diagnosis, or whether instead breast cancer survivors experienced changes in relative earnings, which can be more reliably linked in a causal manner to their cancer. Because the social security data are not helpful in looking at the other outcomes we study—as they include, for example, no information on hours and wages—we also take a couple of other approaches to assessing the influence of selection biases (and other biases) on our estimates. First, we have available a number of control variables, such as wealth, race, age, and education, which are plausibly correlated with survival outcomes as well as labor market behavior. We compare the estimated effects of cancer incidence and survival time with and without these controls. If the inclusion of these controls has much influence on the estimated effects, then an additional role for unobservables seems relatively more likely. On the other hand, if the inclusion of these controls has little or no influence, it seems less likely that other unobserved characteristics would be important. Second, we consider evidence from specific hypotheses regarding selection that might generate spurious effects of breast cancer on the labor market outcomes studied. 4.3. Control variables While the regressor of primary interest is a binary indicator for whether or not women had breast cancer, we are also interested in trying to incorporate information on the duration of survival to date. The literature offers little guidance on how to specify the effects of breast cancer in multivariate analyses predicting labor market outcomes, so we addressed this problem empirically. Our alternatives were to enter breast cancer as either a binary variable indicating the presence or absence of breast cancer or as an ordinal or continuous variable indicating the years since diagnosis. An argument can be made for a continuous specification based on ease of interpretation and parsimony. However, a single-year interval at different times since diagnosis may have unequal impacts on employment (for example) if there is a nonlinear relationship between years since diagnosis and the labor market decisions and outcomes we study. We therefore entered years since diagnosis into our preliminary estimations a number of ways (continuously and as dummy variables) in separate regression equations, and used F-tests to examine the relative contributions of the restricted and unrestricted models. Based on our analysis, we found three primary distinctions among 13 The first wave of the HRS does not collect such information retrospectively. In principle, subsequent waves of the study could be used to construct such data, but the incidence of cancer would likely be too low to generate a sizable treatment group among those initially cancer-free.
764
C.J. Bradley et al. / Journal of Health Economics 21 (2002) 757–779
women: no cancer, cancer diagnosed 2 or fewer years prior to the interview, and cancer diagnosed 3 or more years prior to the interview. We also control for individual characteristics including age, marital status (never married, separated/widowed/divorced, married), race (Hispanic, African-American/non-Hispanic, or Caucasian/Other/non-Hispanic), education (no high school diploma, high school diploma, some college, college degree), health status (poor or fair),14 number of children under age 18, and wealth. Age is specified as a continuous variable plus a dummy variable for women over age 64. The dummy variable controls for Medicare insurance coverage and the typical age at which one retires. We include variables for a spouse having health insurance through his employer (for married women only), region of country (Mid-Atlantic, East North Central, West North Central, South Atlantic, East South Central, West South Central, Mountain, Pacific, and New England), and urban residence. 5. Results 5.1. Descriptive statistics Table 1 presents descriptive statistics for the dependent and independent variables from the HRS. The prevalence estimate of breast cancer is 2.6% (n = 156). Approximately 64% of all women are currently working. They work an average of 38.5 hours per week, have average annual earnings of US$ 18,424, and an average hourly wage of US$ 10.63. Wealth is a highly skewed variable with a mean of US$ 229,531 and median of US$ 98,500. Fifty-five percent of married women’s husbands have health insurance coverage through their employer. Overall, this sample can be described as predominantly Caucasian/other, married, and middle-aged (mean age is 54), and as having a high school education or better and very few children under age 18. Women with breast cancer are similar to the rest of the HRS sample in their demographic characteristics, with the following exceptions. More breast cancer survivors are older, Caucasian/other, married, in poor or fair health, and have achieved higher education, compared to women without breast cancer. These differences are consistent with what we would expect regarding who gets and survives breast cancer. The average time since breast cancer diagnosis is 7 years, with three-quarters of the treatment group having survived for 3 or more years. Thus, we do not expect treatment-related morbidity to affect labor market outcomes for many of these women. Even so, several differences in labor market variables are apparent in this descriptive analysis. Fewer breast cancer survivors work (54% versus 64%). Among those who do work, though, breast cancer survivors work more hours per week and earn more per year. 5.2. Probability of working Table 2 provides the probit model estimates for the probability of working. In the first column breast cancer is captured in a simple dummy variable, and its coefficient is negative 14 Breast cancer coefficients changed slightly when health status was excluded from the equations, indicating that health status absorbs a small portion of the effect of breast cancer.
C.J. Bradley et al. / Journal of Health Economics 21 (2002) 757–779
765
766
C.J. Bradley et al. / Journal of Health Economics 21 (2002) 757–779
C.J. Bradley et al. / Journal of Health Economics 21 (2002) 757–779
767
768
C.J. Bradley et al. / Journal of Health Economics 21 (2002) 757–779
Table 3 Conditional models of average weekly hours worked and hourly wage, Health and Retirement Study 1992 Independent variables
Hours worked
Hours worked
Log hourly wage
Log hourly wage
Breast cancer (yes/no) ≤2 years since diagnosis ≥3 years since diagnosis
3.39∗∗ (1.37) – –
– 0.97 (2.83) 4.05∗∗∗ (1.54)
0.12∗∗ (0.05) – –
– 0.02 (0.10) 0.14∗∗∗ (0.06)
Notes: Sample sizes: columns 1 and 2, N = 3818; and columns 3 and 4, N = 3381. Standard errors are in parentheses. Control variables are the same as those reported in Table 2. ∗∗ Significant at P < 0.05. ∗∗∗ Significant at P < 0.01.
and statistically significant (−0.07, P < 0.10). This estimate implies that at the mean for all other covariates, the estimated probability of women with breast cancer working is seven percentage points lower than for women who do not have breast cancer. In the second column, we differentiate between women diagnosed with breast cancer within 2 years of the interview and those diagnosed 3 or more years prior to survey date. We also test the effect of a spouse holding employer-based health insurance on the probability of working among married women in the HRS sample. The magnitude and significance of the coefficient on breast cancer (column 3) increases slightly to −0.09 (P < 0.10) relative to the estimate for all women in column 1. If a spouse has employer-based insurance, women are 11 percentage points less likely to work. In an alternative specification (not reported), the interaction for having a spouse with employer-based health insurance and breast cancer was not statistically significant. Thus, the hypothesis that breast cancer survivors who have the option of health insurance through their spouses are less likely to work—or equivalently that those without such insurance are more compelled to work—is questionable. 5.3. Hours worked The first two columns of Table 3 show the results for breast cancer from our conditional OLS models for hours worked.15 In the model predicting hours conditional on employment, we find a counterintuitive result, although one that replicates what we saw in the means. Employed breast cancer survivors are estimated to work 3.39 hours (P < 0.05) more a week than other employed women. When we specify years since diagnosis, we see that in the years immediately following diagnosis, women with breast cancer do not work a different number of hours than women without cancer. In subsequent years following diagnosis, however, women with breast cancer work an average of approximately 4 hours additional (P < 0.01) per week. The signs and significance of the remaining coefficients (not reported) were consistent with labor supply theory. 15 We also estimated unconditional OLS models for hours worked. The point estimates of the coefficients on breast cancer were negative but statistically insignificantly different from zero. The lack of any relationship may reflect apparently offsetting influences of breast cancer on employment and on hours conditional on work.
C.J. Bradley et al. / Journal of Health Economics 21 (2002) 757–779
769
5.4. Wages The effect of breast cancer on log hourly wages is also reported in Table 3 (columns 3 and 4) using similar specifications to those for hours worked. This evidence suggests that breast cancer survivors’ wages are higher than the wages of women in the control group. In particular, we find a statistically significant (P < 0.05) coefficient on breast cancer (column 3). Breast cancer survivors earn approximately 13% more per hour than other women.16 In column 4, women who have survived breast cancer for 3 or more years earn approximately 15% more per hour than women without breast cancer, while there is no statistically significant differential for those diagnosed within the past 2 years. 5.5. Earnings In a comparison of mean self-reported annual earnings, working women who had been diagnosed with breast cancer earn on average nearly US$ 2350 more than other women, an advantage of approximately 12% (Table 1). In columns 1 and 2 of Table 4, the estimates indicate that breast cancer survivors earn more than the non-cancer control group when we estimate a log earnings equation (conditional on working). Breast cancer survivors have 21% (P < 0.10) higher earnings than the non-cancer control group. When we further controlled for years since diagnosis, women diagnosed 3 or more years prior to the interview earn approximately 26% more (P < 0.10) than the non-cancer control group. The remaining four columns of the table report estimates of similar specifications with the social security earnings data for the same year, to provide a baseline for the longitudinal analyses discussed later. Paralleling the results with the HRS earnings data, the social security earnings results suggest that working breast cancer survivors earn 39% more than the non-cancer control group. The differential is slightly smaller when we condition on positive earnings in the year prior to diagnosis (or a similar matched year for the control group). In these data, the difference between those diagnosed 2 or fewer years ago and those diagnosed 3 or more years ago is not apparent, so in the ensuing longitudinal analysis we drop this distinction. 5.6. Selection bias assessment with longitudinal social security earnings The cross-sectional results indicate that while breast cancer survivors have lower employment probabilities, those who have survived 3 or more years work more hours, earn higher wages, and have higher earnings, if they are employed. These results are surprising, if one believes that there are any long-lasting debilitating effects of breast cancer. Even if there are not long-lasting effects, the most reasonable expectation might be to find no differentials between the survivors and the non-cancer control group. On the other hand, a positive causal effect is not inconceivable, as some survivors perhaps replenish financial as16 This is based on the suggestion of Halvorsen and Palmquist (1980) for interpreting dummy variables in semilogarithmic equations. In particular, the coefficient of .12 in column 3 implies that wages are higher by {(e.12 − 1) × 100}, or 13%. This calculation is used throughout the paper in describing the results from semilogarithmic specifications.
770
C.J. Bradley et al. / Journal of Health Economics 21 (2002) 757–779
C.J. Bradley et al. / Journal of Health Economics 21 (2002) 757–779
771
sets that may have been depleted while fighting the cancer. We assess the available evidence on whether the estimated relationships described in the previous sections are causal, or are instead driven by selection bias. Selection bias could arise if women who survive breast cancer have more human capital or greater financial resources—differences that may not be fully captured in our control variables—and the unmeasured differences are associated with higher hours, wages, and earnings.17 Lower employment probabilities for breast cancer survivors appear to be more consistent with the hypothesis that we are observing causal effects. Of course the population of survivors could be heterogeneous, with some debilitated by the disease and therefore leaving the labor market, while those who remain are a select group of those most committed to the labor market, or most “robust”. Thus, additional evidence assessing the validity of a causal interpretation of the estimated effects for employment, hours, earnings, and wages is useful. We use social security earnings records that can be linked to the HRS to explore whether the cross-sectional association between breast cancer and earnings—indicating a strong positive impact—appears to be causal. In particular, we estimate three closely related longitudinal specifications to examine the association between breast cancer and earnings, preand post-diagnosis. First, we estimate the association between breast cancer (as a dummy variable) and the log of earnings in the year prior to diagnosis to see if “future breast cancer survivors” were simply higher earners even prior to diagnosis. To obtain comparable data for the control group—for use in all of the longitudinal analyses—we randomly assign an early year to each observation in this group, so as to match the distribution of the year prior to diagnosis for the treatment group of breast cancer survivors. As reported in column 1, Panel 1, of Table 5, we do not find an association between prior earnings and subsequent breast cancer. This indicates that while breast cancer survivors have higher earnings ex post (according to the estimates in Table 4), they do not have higher earnings ex ante, undermining the hypothesis that breast cancer survivors are simply a select group of high earners. For the second specification, we estimate the effect of breast cancer on the log of 1991 social security earnings, controlling for the log of earnings in the year prior to diagnosis or a matched early year for the control group (Table 5, column 2, Panel 2). In an alternative specification (Panel 2), we also control for the number of years between 1991 and the year before diagnosis, with no difference in the results. Breast cancer continues to have a positive and significant effect on earnings that is of the same magnitude (39%, P < 0.01) as in the cross-sectional analysis of social security earnings. For the third specification, reported in column 3 of Table 5, instead of using log earnings as the dependent variable, we use the difference between log 1991 earnings and log earnings in year prior to diagnosis or the matched early year for the control group. Alternative specifications were also estimated including a control for the number of years between earnings observations (Panel 2), and a control for the change in log earnings 2 and 3 years before diagnosis (Panel 3), to allow for different growth rates, and not just different levels, between breast cancer survivors and the control group. Once again our findings are robust; breast can17 The theory of compensating differentials might suggest that women in jobs with higher risk of cancer earn higher wages, although we do not think that this is a major part of the story.
772
C.J. Bradley et al. / Journal of Health Economics 21 (2002) 757–779
Table 5 Longitudinal specifications using social security earnings records Independent variables
Panel 1 Breast cancer (yes/no) Log earnings, year before diagnosis Panel 2 Breast cancer (yes/no) Log earnings, year before diagnosis Number of years between year before diagnosis and 1991 Panel 3 Breast cancer (yes/no) (Log earnings 2 years before diagnosis)− (Log earnings 3 years before diagnosis) Number of years between year before diagnosis and 1991
Log earnings year before diagnosis
Log 1991 earnings
(Log 1991 earnings)− (Log earnings of year before diagnosis)
−0.01 (0.17) –
0.33∗∗∗ (0.11) 0.47∗∗∗ (0.03)
0.33∗ (0.18) –
– – –
0.34∗∗∗ (0.11) 0.48∗∗∗ (0.03) 0.02∗∗∗ (0.004)
0.34∗ (0.18) – 0.03∗∗∗ (0.01)
– –
– –
–
–
0.35∗ (0.18) −0.03∗ (0.01) 0.03∗∗∗ (0.01)
Notes: N = 2432. Sample is restricted to women with positive earnings in “year before diagnosis” and 1991. For the control group, “year before diagnosis” is an observation from an earlier year drawn from the distribution of years prior to diagnosis for breast cancer survivors. Standard errors are in parentheses. Control variables are the same as those reported in Table 4. While earlier observations on these variables are not available, most are relatively fixed over the later part of the life cycle, although they may still enter the differenced specifications in column 3 if they are associated with differences in earnings growth. ∗ Significant at P < 0.10. ∗∗∗ P < 0.01.
cer survivors have significantly greater growth in earnings between the year prior to breast cancer diagnosis and 1991 than do women in the control group over a comparable period. Thus, the longitudinal evidence supports the conclusion that breast cancer survivorship is associated with an increase in earnings for those women who continue to work. 5.7. Further assessment of selection bias Finally, we turn to some additional evidence assessing a causal interpretation of our broader set of findings regarding breast cancer survival and labor market outcomes. First, we report analyses that compare our findings reported earlier for the probability of work, and for hours, wages, and earnings, to estimates of models where we do not control for wealth, race, age, and education. These variables have been identified as associated with surviving breast cancer and with labor market outcomes. If the exclusion of these controls does not substantively influence the estimated effects of breast cancer survival, then the argument for selection into the treatment group on unobservables is weakened. The results from this analysis are shown in Table 6. By comparing these estimates to those in the earlier tables, the reader can see that the estimated effects on employment, hours, wages, and earnings are virtually unchanged in this table.18 In contrast, if we were “worsening” the 18
The conclusion regarding earnings was also true if we used social security earnings as the dependent variable.
C.J. Bradley et al. / Journal of Health Economics 21 (2002) 757–779
773
774
C.J. Bradley et al. / Journal of Health Economics 21 (2002) 757–779
C.J. Bradley et al. / Journal of Health Economics 21 (2002) 757–779
775
776
C.J. Bradley et al. / Journal of Health Economics 21 (2002) 757–779
selection on unobservables in this table, we would expect the effects to be larger in Table 6. This provides suggestive evidence that unobservable characteristics associated with breast cancer are not influencing the labor market outcomes we observe, thus providing further support for a causal interpretation of our findings. Our next set of analyses pursue more directly the hypothesis that, among breast cancer survivors, those who exhibit higher hours of work, earnings, and wages ex post were simply more attached to the labor market and were higher-wage workers ex ante. The HRS includes job tenure information if the respondent is currently working. If the respondent is not working, job tenure information is collected for the respondent’s most recent job, unless the respondent never worked for more than a few months or last worked prior to 1972. Thus, there is a sample for which we can include job tenure as a proxy for labor market commitment or attachment. This analysis is reported in Table 7. Because there are some observations for which tenure is unavailable, we first report estimates of the earlier models using the smaller sample for which the tenure information is available (in the upper panel). Otherwise, the estimated specifications including tenure would not be comparable. Then, in the lower panel, we re-estimate the models including the tenure variable. The table reveals that the estimated differentials associated with breast cancer are largely unaffected by the inclusion of the tenure control, although they fall slightly. In our view, the fact that these relationships are little changed by including a variable strongly related to past labor market commitment also helps to undermine a selection explanation of our findings. Our final analysis tests a specific version of a selection story, namely the hypothesis that the hours differential we find arises because women working part-time are not as attached to the labor force as full-time workers. When these women are presented with a breast cancer diagnosis, these women may quit their jobs, whereas women working full-time with a breast cancer diagnosis continue to work. This could also explain the higher earnings of breast cancer survivors, as well as their higher wages (since full-time workers earn a wage premium). If this is true, we should find that breast cancer survivors are relatively more likely to work full-time, and that the hours differential should disappear when we subdivide the sample into full- and part-time workers. The results of this analysis are shown in Table 8. Columns 1 and 2 show no statistically significant differential in the probability of being employed 35 or more hours per week associated with breast cancer. In columns 3 and 4, we do not find that breast cancer has an effect on the number of hours worked per week for women employed full-time. Furthermore, in columns 5 and 6 we show a positive effect for women working less than 35 hours per week and diagnosed with breast cancer. Thus, the data are not consistent with the hypothesis that women employed part-time tend to leave their jobs after a breast cancer diagnosis. Rather, among survivors part-time women appear to work more hours.
6. Discussion Our finding that breast cancer has a negative effect on working is consistent with what one would expect given the disease and population affected, and our sensitivity analysis indicates that the coefficients on the breast cancer variables for this labor market outcome are relatively robust. However, breast cancer’s positive effect on hours worked, wages, and
C.J. Bradley et al. / Journal of Health Economics 21 (2002) 757–779
777
earnings is inconsistent with standard views of the natural course of the disease (i.e. that it should either reduce the number of hours worked and productivity because of recurrences or health effects, or at best be neutral if the cancer remains in remission and has no lasting debilitating effects), which is why we have delved at length into the question of selection bias. We believe that the various analyses of this issue are complementary, and the evidence generally suggests that our findings are not driven by selection bias. We might think that the best-case scenario is that the morbidity associated with breast cancer is negligible for at least some women, so that for these women, who are more likely to continue to work, there are no adverse consequences in terms of labor market outcomes or performance. There are several studies indicating that the morbidity associated with certain types and stages of breast cancer and its treatment may not interfere with work (e.g. Ganz et al., 1996; Satariano and DeLorenze, 1996). But this explanation is only consistent with no effects, not positive effects. One potential explanation of positive effects is that women may be unwilling to reduce number of hours worked or to exit the labor force for fear of not getting another job with health insurance benefits. Relatively few part-time jobs provide health insurance benefits (Olson, 1998). However, our analysis of the interaction between spousal insurance coverage and breast cancer and their combined effect on hours worked does not support this hypothesis, although we believe this question merits further research. Even in the absence of ideal data, our analysis improves upon what is known in the existing literature. Our study, controlling for a range of correlates, provides descriptive information on differences in labor market outcomes between cancer survivors and a non-cancer control group. We also provide objective measures of individuals’ economic circumstances after a cancer diagnosis. Finally, with respect to one outcome—earnings—we are able to observe the actual longitudinal developments a cancer survivor experiences, and to compare this with a control group. We are currently collecting longitudinal data to address this issue more fully with respect to a broader array of labor market outcomes and a wider set of cancers, and so our future research should lend further insight into the causal effects of cancer on survivors’ labor market experiences. The American Cancer Society estimates that there will be 175,000 new cases of invasive breast cancer this year among women in the United States. In general, the high incidence of certain cancer types among working age people leads to greater productivity costs for cancer compared to other categories of disease. An understanding of the labor market implications of a cancer diagnosis and treatment is particularly relevant as cancer screening (including prostate, colon, and lung, in addition to breast) is routinely applied to a working age population where these cancers may have otherwise gone undetected for some time.
Acknowledgements We are grateful to Charles W. Given, Margaret Holmes-Rovner, Daiji Kawaguchi, Donna Edwards Neumark, and two anonymous reviewers for helpful comments. Bradley’s research was supported by an NCI grant, RO1-CA86045-01A1. Neumark’s research was partially supported by NIA grant K01-AG00589.
778
C.J. Bradley et al. / Journal of Health Economics 21 (2002) 757–779
References Adams, S.J., 2001. Employer-provided health insurance and job change: new evidence of job lock. University of Michigan, Ann Arbor, Michigan, unpublished. American Cancer Society, 1999. Cancer Facts and Figures. American Cancer Society, Atlanta, GA. American Cancer Society, 2001. Cancer Facts and Figures. American Cancer Society, Atlanta, GA. Benham, L., Benham, A., 1982. Employment, earnings, and psychiatric disorders. In: Fuchs, V. (Ed.), Economic Aspects of Health. University of Chicago Press, Chicago, Illinois, pp. 203–220. Berkman, B.J., Sampson, S.E., 1993. Psychosocial effects of cancer economics on patients and their families. Cancer 72 (Suppl.), 2846–2849. Berry, D., 1993. Return-to-work experiences of people with cancer. Oncology Nursing Forum 20 (6), 905–911. Bradley, C.J., Bednarek, H.L., Neumark, D. Breast cancer and women’s labor supply. Health Services Research, in press. Buchmueller, T.C., Valletta, R.G., 1999. The effect of health insurance on married female labor supply. Journal of Human Resources 34, 42–70. Carter, B., 1994. Surviving breast cancer: a problematic work re-entry. Cancer Practice 2 (2), 135–140. Currie, J., Madrian, B.C., 1999. Health, health insurance and the labor market. In: Ashenfelter, O., Card, D., (Eds.), Handbook of Labor Economics, Vol. 3. Elsevier, Amsterdam, pp. 3309–3416. Ettner, S.L., Frank, R.G., Kessler, R., 1997. The impact of psychiatric disorders on labor market outcomes. Industrial and Labor Relations Review 51 (1), 64–81. Famulari, M., 1992. The effects of a disability on labor market performance: the case of epilepsy. Southern Economic Journal 58 (4), 1072–1087. Fow, N.R., 1996. Cancer rehabilitation: an investment in survivorship. As more people survive the disease, focus shifts on improving quality of life. Rehabilitation Management 9 (3), 48–53. Friedland, R.B., 1996. People with disabilities: access to health care and related benefits. In: Mashaw, J.L. (Ed.), Disability, Work and Cash Benefits. W.E. Upjohn Institute for Employment Research, Kalamazoo, Michigan, pp. 357–388. Ganz, P.A., Coscarelli, A., Fred, C., Kahn, B., Polinsky, M.L., Petersen, L., 1996. Breast cancer survivors: psychosocial concerns and quality of life. Breast Cancer Research and Treatment 38 (2), 183–199. Greenwald, H.P., Dirks, S.J., Borgatta, E.F., McCorkle, R., Nevitt, M.C., Yelin, E.H., 1989. Work disability among cancer patients. Social Science and Medicine 29 (11), 1253–1259. Grossman, M., 1972. On the concept of health capital and the demand for health. Journal of Political Economy 80 (2), 223–255. Gruber, J., Madrian, B.C., 1994. Health insurance and job mobility: the effects of public policy on job-lock. Industrial and Labor Relations Review 48 (1), 86–102. Halvorsen, R., Palmquist, R., 1980. The interpretation of dummy variables in semilogarithmic equations. American Economic Review 70 (3), 474–475. Heckman, J.J., 1979. Shadow prices, market wages, and labor supply. Econometrica 42 (4), 679–694. Kapur, K., 1998. The impact of health on job mobility: a measure of job-lock. Industrial and Labor Relations Review 51 (2), 282–298. Kornblith, A.B., 1998. Psychosocial adaptation of cancer survivors. In: Holland, J.C. (Ed.), Psycho-Oncology. Oxford University Press, New York, pp. 223–241. Madrian, B.C., 1994. Employment-based health insurance and job mobility: is there evidence of job lock? Quarterly Journal of Economics 109 (1), 27–54. Moffitt, R., Wolfe, B., 1992. The effect of the Medicaid program on welfare participation and labor supply. The Review of Economics and Statistics 74 (4), 615–626. Monheit, A.C., Cooper, P.F., 1994. Health insurance and job mobility: theory and evidence. Industrial and Labor Relations Review 48 (1), 68–85. Olsen, O., Gotzsche, P., 2001. Cochrane review on screening for breast cancer with mammography. Lancet 358 (9290), 1284–1285. Olsen, R.J., 1980. A least squares correction for selectivity bias. Econometrica 48 (7), 1815–1820. Olson, C., 1998. A comparison of parametric and semiparametric estimates of the effect of spousal health insurance coverage on weekly hours worked by wives. Journal of Applied Econometrics 14, 543–565.
C.J. Bradley et al. / Journal of Health Economics 21 (2002) 757–779
779
Quigley, K.M., 1989. The adult cancer survivor: psychosocial consequences of cure. Seminar in Oncology Nursing 5, 63–69. Razavi, D., Delvaux, N., Bredart, A., Autier, P., Bron, D., Debusscher, L., Stryckmans, P., 1993. Professional rehabilitation of lymphoma patients: a study of psychosocial factors associated with return to work. Support Care Cancer 1, 276–278. Satariano, W.A., DeLorenze, G.N., 1996. The likelihood of returning to work after breast cancer. Public Health Reports 111 (3), 236–241. Schagen, S.B., Frits, S.A.M., van Dam, M., Muller, J., Boogerd, W., Lindeboom, J., Bruning, P.F., 1999. Cognitive deficits after postoperative adjuvant chemotherapy for breast carcinoma. Cancer 85 (3), 640–650.
Breast cancer survival, work, and earnings Cathy J. Bradley a,∗ , Heather L. Bednarek b , David Neumark c a
Department of Medicine, College of Human Medicine, B212 Clinical Center, Michigan State University, East Lansing, MI 48824, USA b Department of Economics, St. Louis University, Saint Louis, MO 63108, USA c Department of Economics, Michigan State University, East Lansing, MI 48824, USA Received 1 May 2001; accepted 1 February 2002
Abstract Relying on data from the Health and Retirement Study (HRS) linked to longitudinal social security earnings data, we examine differences between breast cancer survivors and a non-cancer control group in employment, hours worked, wages, and earnings. Overall, breast cancer has a negative impact on employment. However, among survivors who work, hours of work, wages, and earnings are higher compared to women in the control group. We explore possible biases underlying these estimates, focusing on selection, but cannot rule out a causal interpretation. Our research points to heterogeneous labor market responses to breast cancer, and shows that breast cancer does not appear to be debilitating for women who remain in the work force. © 2002 Elsevier Science B.V. All rights reserved. JEL classification: I1; E61; J21 Keywords: Breast cancer; Employment; Labor market effects; Hours worked; Earnings
1. Introduction In recent years, improved detection methods for breast cancer have led to the treatment and survival of a younger population of women more likely to be at working ages, making an inquiry into the impact of breast cancer on labor market outcomes particularly relevant. Between 1983 and 1993, in situ breast cancer rates increased from 2.3 to 6.2 per 100,000 among women under age 50, largely reflecting an increase in the use of mammography (American Cancer Society, 1999). Treatment has improved as well, leading to the largest short-term decline in over 40 years in breast cancer mortality (American Cancer Society, ∗ Corresponding author. Tel.: +1-517-432-3405; fax: +1-517-432-9471. E-mail address: [email protected] (C.J. Bradley).
0167-6296/02/$ – see front matter © 2002 Elsevier Science B.V. All rights reserved. PII: S 0 1 6 7 - 6 2 9 6 ( 0 2 ) 0 0 0 5 9 - 0
758
C.J. Bradley et al. / Journal of Health Economics 21 (2002) 757–779
2001). Breast cancer research has focused on detection and treatment and, to a lesser extent, survivors’ quality-of-life. However, now that 5-year survival is expected for most women diagnosed in the early stages of breast cancer, attention should also be given to economic measures of the consequences of surviving breast cancer, as part of a broader effort to understand the quality-of-life implications of the disease, and how labor market agents react to cancer survivorship. We focus on breast cancer in this paper because we believe that it is in some important respects unique and worthy of study as a single disease. Breast cancer mostly affects women, and screening for breast cancer (i.e. mammography) is routinely applied to working age women, yet the long-term benefits of screening and treatment of early stage cancers are unknown (Olsen and Gotzsche, 2001), making an inquiry into breast cancer’s effect on productivity particularly relevant. Our view is that other cancers do not share breast cancer’s characteristics in terms of its screening, treatment, and prognosis, and that few other diseases have the same emotional impact as “cancer.” In this paper, we use the first wave of the Health and Retirement Study (HRS)—in some cases linked to longitudinal social security earnings data—in an attempt to understand how breast cancer influences labor market decisions and outcomes. As cancer screening is more routinely applied to a working age population, more cancers are likely to be detected in early stages that may have otherwise gone undetected until a later time. Therefore, this research is particularly relevant as it both fills a gap in the literature regarding labor market consequences of cancer and provides information on how early detection of disease may affect labor market outcomes.
2. Illness and labor market outcomes Intuitively, poor health would seem to have a negative impact on labor supply and productivity. A more formal way of thinking about this impact is that under the assumption of utility maximization, the supply of work hours H (the difference between a time endowment T and the demand for leisure hours L) is determined by tastes, prices, and endowments (e.g. of wealth). Poor health can affect labor supply by diminishing tastes for work and thereby raising the marginal value of leisure time, reducing productivity, and, put simply, by stealing time away from work for health maintenance (Grossman, 1972). Formally, this could be captured by incorporating a health production function into a labor supply model. Thinking beyond this basic model, the potential for a change in health insurance status may be an important consideration in the decision to exit the labor force, given that breast cancer treatment is expensive and the potential for future medical expenses is great. Under these circumstances, an individual’s tolerance for financial and personal risk becomes important. The link between health insurance and employment, particularly in the United States, may create unusual incentives regarding the decision to work (Currie and Madrian, 1999). For people with medical conditions who need specialized (and often expensive) care, obtaining coverage is uncertain (Friedland, 1996). Thus, some people may be trapped in a job for fear of losing insurance and others may have disincentives for leaving public programs and seeking employment since having a job can mean losing coverage (Adams, 2001; Madrian, 1994). Therefore, the counterintuitive notion that in the face of illness labor
C.J. Bradley et al. / Journal of Health Economics 21 (2002) 757–779
759
supply may actually increase or remain unchanged, through the influence of health insurance and its accompanying incentives, is plausible. This tension between health conditions and their negative influence on labor supply, and the demand for health insurance and its potentially positive influence on labor supply, is reflected in the existing empirical research. In empirical studies, the relationship between poor health or disability, health insurance, and labor supply appears unresolved (for an overview see Currie and Madrian, 1999). Psychiatric disorders, for example, have been shown to significantly reduce employment, work hours, and income (Ettner et al., 1997; Benham and Benham, 1982). Severity of illness also plays an important role in the decision to reduce participation in the labor force (Famulari, 1992). Conversely, there is some evidence consistent with the demand for health insurance increasing or maintaining labor supply in response to illness. An ill person— particularly someone with a chronic condition requiring specialized care—is motivated to remain at work to maintain health insurance coverage. Empirical studies (e.g. Gruber and Madrian, 1994; Madrian, 1994; Monheit and Cooper, 1994; Kapur, 1998) supporting this position largely address the potential for job lock that may decrease workers’ mobility, rather than directly addressing workers’ willingness to exit the labor force or curtail hours. Cancer-specific empirical studies have pointed to some key questions in understanding the direction of influence cancer is likely to have on labor market decisions and outcomes, but they are far from decisive, and do not focus in a very detailed manner on the empirical estimation of the effects of cancer on these decisions and outcomes. In general, studies (e.g. Quigley, 1989; Fow, 1996; Kornblith, 1998; Razavi et al., 1993; Carter, 1994; Berkman and Sampson, 1993) have focused on survivors’ subjective impressions of the impact of cancer on their lives. These studies suggest negative factors that can reduce employment, including physical disability (e.g. limitations in upper body strength; Fow, 1996), memory loss (Schagen et al., 1999), lack of control over schedules, need for transportation, and type of work performed (Greenwald et al., 1989; Satariano and DeLorenze, 1996), and in some cases discrimination on the part of employers (Carter, 1994; Berry, 1993). Other studies of employment following breast cancer suggest an alternative view. For example, in a preliminary study, we found that breast cancer survivors work more hours per week than a non-cancer control group (Bradley et al., in press). In a study of patients 2 and 3 years after their primary treatment, Ganz et al. (1996) found that breast cancer survivors function at a high level. Sixty-five percent of survivors (n = 139) were either working for pay or volunteering their services. The mean numbers of hours worked were 34.4 and 33.2 hours per week among women who were 2 and 3 years post-treatment, respectively. This study concluded that women continue to work and perform their usual roles after treatment for breast cancer. Satariano and DeLorenze (1996) had similar findings. In a study of 300 women working at the time of their breast cancer diagnosis, 71% returned to work within 3 months after diagnosis. The implication of these studies is that the overall impact of breast cancer on employment status is minor. However, these studies are not based on national samples, nor do they include a control group. Our analysis contributes to research on this topic along both of these dimensions. In addition, unlike much of the previous research, we control for a number of correlates that can influence the probability of employment, decisions regarding hours of work, and earnings, including individual characteristics (in some analyses unobserved characteristics) and information on health insurance coverage.
760
C.J. Bradley et al. / Journal of Health Economics 21 (2002) 757–779
A fundamental problem presented by our data set is that we analyze breast cancer survivors and a control group. This raises the possibility that differences we find between these two groups are attributable to non-random (conditional on the observables) differences between survivors and non-survivors—based on “selection” into the sample of survivors— rather than to actual effects of breast cancer. We try to address the issue as best as possible based on the observable information we have. In addition, though, when we look at earnings—perhaps the best summary measure of labor market outcomes—we are able to exploit a unique feature of our data set that allows us to link to longitudinal administrative data on earnings of breast cancer survivors and the control group. With these data, we use the pre-diagnosis observations on the survivors, and examine differences relative to earlier observations on the control group, to control for possible selection bias. 3. Data Our main data set is drawn from the first wave (1992) of the HRS. The HRS is designed to answer research questions regarding the relationship between health, income, wealth, job decisions, and retirement. The cohort is aged 51–61.1 In most cases, if there are two members of the household (e.g. husband and wife) each is interviewed regarding his/her own employment history, retirement, health, and demographic characteristics. The more financially knowledgeable respondent in the household is interviewed regarding housing, net worth, income, and health insurance status for the members of the household included in the survey.2 Only one member of the household must be in the age range of the sample frame; therefore spouses may be interviewed even if they are out of the specified age range, and thus our sample departs from national representativeness for these respondents. For our non-cancer control group, we selected women who answered “no” to the question, “Has a doctor ever told you that you have cancer or a malignant tumor of any kind?” Given that a cancer diagnosis is a life-altering event, recall bias is not likely to be a problem in the data. Those women that said “yes” to the above question and subsequently indicated that their diagnosis was breast cancer constitute the treatment group. We also consider information available in the HRS on the number of years since breast cancer diagnosis, which is likely to be associated with morbidity and the severity of any side effects from treatment, and therefore to have potential labor market consequences. Women in our sample were diagnosed an average of 7 years (standard deviation (S.D.) = 6, range = 1–32) prior to the interview. The HRS contains 6708 women on whom we have information regarding labor force participation and hours worked. We excluded women who were insured by Medicaid (n = 319) because the program has been shown to constrain labor supply, particularly for families with high expected medical expenditures (Moffitt and Wolfe, 1992). The non-Medicaid restriction coupled with missing data3 led to a final sample size of 5974 women, 156 of whom reported having breast cancer. In addition, to carry out the longitudinal analysis of 1
There are follow-up interviews at 2-year increments, which are not used in this study. If a person was single, he or she was designated the primary respondent. In a married pair or partnership, the person most knowledgeable of household financial matters was designated the primary respondent. 3 Exclusions and missing data were distributed as follows: Medicaid (319), age (120), type of cancer (282), race (13). 2
C.J. Bradley et al. / Journal of Health Economics 21 (2002) 757–779
761
earnings discussed earlier we obtained the restricted social security annual earnings file (covering years 1951–1991) for HRS respondents. Social security earnings were reported for 4527 women; 2936 (68 breast cancer survivors) of them worked in 1991.
4. Empirical approach 4.1. Specifications The outcomes of interest are employment status, usual hours of weekly work, wages, and earnings. Each of these can be expressed as a function of the incidence of breast cancer (BCA), the duration of survival to date (S), exogenous variables (X), health insurance (HI), and random or unobserved influences (ε). Generically, we write the employment equation as: Ei∗ = f (BCAi , Si , Xi , HIi , εi )
(1)
Ei∗
where represents a latent variable for the propensity for employment. We estimate the employment equation by assuming a linear additive form for Eq. (1),4 Ei∗ = αE + βE BCAi + Xi γE + δE HIi + εi
(2)
We then define employment status as a binary variable (Ei ) that equals one if the respondent reports positive hours worked and zero otherwise, and estimate Eq. (2) as a probit model. For ease of interpretation, the probit estimates are translated into derivatives of the probability of working with respect to the independent variables.5 Health insurance may affect labor supply as well as mediate the effects of breast cancer. It is available in large part, for the women in our sample, if they are employed or married with an employed spouse. It is highly likely that the presence of employer-based health insurance is endogenous to a model estimating employment, and so we approach the estimation of the effect of health insurance, for married women, using information on whether her spouse has health insurance through his employer (Buchmueller and Valletta, 1999; Olson, 1998). Consequently, when we estimate Eq. (2) including information on spousal health insurance, we use a sample of married women only. In our estimation of the effects of breast cancer on hours (H), wages (W), and earnings (Y), we assume that the same variables that affect employment also affect these dependent variables. We focus on the estimates of the conditional functions for: Hi = αH + βH BCAi + Xi γH + ηi ,
if Ei = 1,
Wi = αW + βW BCAi + Xi γW + µi , Yi = αY + βY BCAi + Xi γY + νi , 4
if Ei = 1,
if Ei = 1,
(3) (4) (5)
Note that we have omitted Si for now; later, we introduce information on length of survivorship. The “derivatives” are computed as the difference in probabilities as the dummy variable takes on the values zero and one, with the other variables at the sample means. The standard errors reported are those making the t-statistics for these derivatives the same as those on the original coefficients, allowing the reader to easily assess the statistical significance of the raw coefficients. 5
762
C.J. Bradley et al. / Journal of Health Economics 21 (2002) 757–779
recovering estimates of the linear projections of Hi , Wi , or Yi on the independent variables for those working.6,7,8,9 Hours are defined as average hours worked (H) per week while earnings (Y) are based on the respondent’s reported annual earnings in 1991 (and in some cases social security reported annual earnings). Wages (W) were reported on an hourly basis by many HRS respondents (n = 1956). For respondents who reported their earnings in weekly, bi-weekly, monthly, or annual increments, we computed wages by dividing their reported earnings by their average hours worked during the specified time frame. For example, if the respondent worked an average of 40 hours per week and earnings were reported bi-weekly, we divided bi-weekly earnings by 80.10 4.2. Econometric issues In order for a breast cancer survivor to appear in our data set, she must have survived breast cancer for some time. In the HRS sample, 76% of the patients had survived 3 or more years. It has been argued that labor market decisions and health should be jointly considered (Grossman, 1972; Olson, 1998). This is a potentially futile challenge given our data, since we know very little about respondents’ risk factors for breast cancer (e.g. family history, nulliparity, estrogen use, high fat diets, and environmental risks). Thus, treating the incidence of breast cancer as essentially random, other than a few known exogenous predictors out of individuals’ control such as age and race,11 seems a reasonable approximation to reality. Likewise, survival is threatened by cancer recurrences and metastases-events that may also be related to unobservable variables such as breast cancer stage at diagnosis, histology, and other comorbidities. A causal interpretation of the estimates of the equations described in the previous section requires that survival be unrelated to the unobservables.12 To fully address this issue, we would require longitudinal data on women before and after the onset of breast cancer (along with clinical information on their cancer). In the absence of such data and under weaker assumptions (that the unobserved residuals are fixed over time for individuals), we can recover causal estimates by comparing changes in behavior for 6
We assume a linear functional form for hours worked and a log-linear functional form for wages and earnings. We estimated Eqs. (3) and (5) for married women, controlling for whether her spouse had health insurance through an employer. However, the results did not qualitatively change. We do not estimate Eq. (4) in this manner, as spouse’s health insurance is unlikely to independently affect productivity. 8 Such regressions generally provide biased estimates of the slopes of the unconditional mean function (unless the conditional expectations of the error terms are uncorrelated with the regressors), but still are informative about how the exogenous variables are related to hours of work (wages and earnings) for those at work. The selection–correction model (Heckman, 1979) is inappropriate since there are no natural candidates for variables that could be excluded a priori from the second-stage equation, and identification would rely on functional form restrictions (e.g. linearity) or distributional assumptions (see Olsen, 1980). 9 In all cases we also estimated Tobit models, which are consistent if the normality assumption holds but not otherwise; throughout the empirical analysis, the results were not sensitive to the choice of estimator. Results are available upon request. 10 Extreme outliers in reported or computed wages were deleted, with no substantive impact on the results. 11 Caucasian women are slightly more likely to develop but also to survive breast cancer compared to other women. 12 Similarly, consistent estimation of the conditional mean specifications including S (the number of years of survival thus far) requires that this survival duration also be unrelated to the unobservables. 7
C.J. Bradley et al. / Journal of Health Economics 21 (2002) 757–779
763
cancer patients (pre- and post-diagnosis/survival) to changes in behavior for the non-cancer control group over a similar period. Such data are by and large absent in the HRS.13 However, one exception is the linked social security annual earnings records available for each of the up to 40 years prior to the interview, which can be obtained from the administrators of the HRS through special arrangements designed to protect respondents’ confidentiality. Therefore, we first address the issue of causality by exploiting the information in the social security data on earnings in the years prior to diagnosis, to conduct longitudinal analyses. This allows us to examine earnings pre- and post-diagnosis for breast cancer survivors and at comparable times for the control group, to see if post-diagnosis differentials between breast cancer survivors and the control group also existed prior to diagnosis, or whether instead breast cancer survivors experienced changes in relative earnings, which can be more reliably linked in a causal manner to their cancer. Because the social security data are not helpful in looking at the other outcomes we study—as they include, for example, no information on hours and wages—we also take a couple of other approaches to assessing the influence of selection biases (and other biases) on our estimates. First, we have available a number of control variables, such as wealth, race, age, and education, which are plausibly correlated with survival outcomes as well as labor market behavior. We compare the estimated effects of cancer incidence and survival time with and without these controls. If the inclusion of these controls has much influence on the estimated effects, then an additional role for unobservables seems relatively more likely. On the other hand, if the inclusion of these controls has little or no influence, it seems less likely that other unobserved characteristics would be important. Second, we consider evidence from specific hypotheses regarding selection that might generate spurious effects of breast cancer on the labor market outcomes studied. 4.3. Control variables While the regressor of primary interest is a binary indicator for whether or not women had breast cancer, we are also interested in trying to incorporate information on the duration of survival to date. The literature offers little guidance on how to specify the effects of breast cancer in multivariate analyses predicting labor market outcomes, so we addressed this problem empirically. Our alternatives were to enter breast cancer as either a binary variable indicating the presence or absence of breast cancer or as an ordinal or continuous variable indicating the years since diagnosis. An argument can be made for a continuous specification based on ease of interpretation and parsimony. However, a single-year interval at different times since diagnosis may have unequal impacts on employment (for example) if there is a nonlinear relationship between years since diagnosis and the labor market decisions and outcomes we study. We therefore entered years since diagnosis into our preliminary estimations a number of ways (continuously and as dummy variables) in separate regression equations, and used F-tests to examine the relative contributions of the restricted and unrestricted models. Based on our analysis, we found three primary distinctions among 13 The first wave of the HRS does not collect such information retrospectively. In principle, subsequent waves of the study could be used to construct such data, but the incidence of cancer would likely be too low to generate a sizable treatment group among those initially cancer-free.
764
C.J. Bradley et al. / Journal of Health Economics 21 (2002) 757–779
women: no cancer, cancer diagnosed 2 or fewer years prior to the interview, and cancer diagnosed 3 or more years prior to the interview. We also control for individual characteristics including age, marital status (never married, separated/widowed/divorced, married), race (Hispanic, African-American/non-Hispanic, or Caucasian/Other/non-Hispanic), education (no high school diploma, high school diploma, some college, college degree), health status (poor or fair),14 number of children under age 18, and wealth. Age is specified as a continuous variable plus a dummy variable for women over age 64. The dummy variable controls for Medicare insurance coverage and the typical age at which one retires. We include variables for a spouse having health insurance through his employer (for married women only), region of country (Mid-Atlantic, East North Central, West North Central, South Atlantic, East South Central, West South Central, Mountain, Pacific, and New England), and urban residence. 5. Results 5.1. Descriptive statistics Table 1 presents descriptive statistics for the dependent and independent variables from the HRS. The prevalence estimate of breast cancer is 2.6% (n = 156). Approximately 64% of all women are currently working. They work an average of 38.5 hours per week, have average annual earnings of US$ 18,424, and an average hourly wage of US$ 10.63. Wealth is a highly skewed variable with a mean of US$ 229,531 and median of US$ 98,500. Fifty-five percent of married women’s husbands have health insurance coverage through their employer. Overall, this sample can be described as predominantly Caucasian/other, married, and middle-aged (mean age is 54), and as having a high school education or better and very few children under age 18. Women with breast cancer are similar to the rest of the HRS sample in their demographic characteristics, with the following exceptions. More breast cancer survivors are older, Caucasian/other, married, in poor or fair health, and have achieved higher education, compared to women without breast cancer. These differences are consistent with what we would expect regarding who gets and survives breast cancer. The average time since breast cancer diagnosis is 7 years, with three-quarters of the treatment group having survived for 3 or more years. Thus, we do not expect treatment-related morbidity to affect labor market outcomes for many of these women. Even so, several differences in labor market variables are apparent in this descriptive analysis. Fewer breast cancer survivors work (54% versus 64%). Among those who do work, though, breast cancer survivors work more hours per week and earn more per year. 5.2. Probability of working Table 2 provides the probit model estimates for the probability of working. In the first column breast cancer is captured in a simple dummy variable, and its coefficient is negative 14 Breast cancer coefficients changed slightly when health status was excluded from the equations, indicating that health status absorbs a small portion of the effect of breast cancer.
C.J. Bradley et al. / Journal of Health Economics 21 (2002) 757–779
765
766
C.J. Bradley et al. / Journal of Health Economics 21 (2002) 757–779
C.J. Bradley et al. / Journal of Health Economics 21 (2002) 757–779
767
768
C.J. Bradley et al. / Journal of Health Economics 21 (2002) 757–779
Table 3 Conditional models of average weekly hours worked and hourly wage, Health and Retirement Study 1992 Independent variables
Hours worked
Hours worked
Log hourly wage
Log hourly wage
Breast cancer (yes/no) ≤2 years since diagnosis ≥3 years since diagnosis
3.39∗∗ (1.37) – –
– 0.97 (2.83) 4.05∗∗∗ (1.54)
0.12∗∗ (0.05) – –
– 0.02 (0.10) 0.14∗∗∗ (0.06)
Notes: Sample sizes: columns 1 and 2, N = 3818; and columns 3 and 4, N = 3381. Standard errors are in parentheses. Control variables are the same as those reported in Table 2. ∗∗ Significant at P < 0.05. ∗∗∗ Significant at P < 0.01.
and statistically significant (−0.07, P < 0.10). This estimate implies that at the mean for all other covariates, the estimated probability of women with breast cancer working is seven percentage points lower than for women who do not have breast cancer. In the second column, we differentiate between women diagnosed with breast cancer within 2 years of the interview and those diagnosed 3 or more years prior to survey date. We also test the effect of a spouse holding employer-based health insurance on the probability of working among married women in the HRS sample. The magnitude and significance of the coefficient on breast cancer (column 3) increases slightly to −0.09 (P < 0.10) relative to the estimate for all women in column 1. If a spouse has employer-based insurance, women are 11 percentage points less likely to work. In an alternative specification (not reported), the interaction for having a spouse with employer-based health insurance and breast cancer was not statistically significant. Thus, the hypothesis that breast cancer survivors who have the option of health insurance through their spouses are less likely to work—or equivalently that those without such insurance are more compelled to work—is questionable. 5.3. Hours worked The first two columns of Table 3 show the results for breast cancer from our conditional OLS models for hours worked.15 In the model predicting hours conditional on employment, we find a counterintuitive result, although one that replicates what we saw in the means. Employed breast cancer survivors are estimated to work 3.39 hours (P < 0.05) more a week than other employed women. When we specify years since diagnosis, we see that in the years immediately following diagnosis, women with breast cancer do not work a different number of hours than women without cancer. In subsequent years following diagnosis, however, women with breast cancer work an average of approximately 4 hours additional (P < 0.01) per week. The signs and significance of the remaining coefficients (not reported) were consistent with labor supply theory. 15 We also estimated unconditional OLS models for hours worked. The point estimates of the coefficients on breast cancer were negative but statistically insignificantly different from zero. The lack of any relationship may reflect apparently offsetting influences of breast cancer on employment and on hours conditional on work.
C.J. Bradley et al. / Journal of Health Economics 21 (2002) 757–779
769
5.4. Wages The effect of breast cancer on log hourly wages is also reported in Table 3 (columns 3 and 4) using similar specifications to those for hours worked. This evidence suggests that breast cancer survivors’ wages are higher than the wages of women in the control group. In particular, we find a statistically significant (P < 0.05) coefficient on breast cancer (column 3). Breast cancer survivors earn approximately 13% more per hour than other women.16 In column 4, women who have survived breast cancer for 3 or more years earn approximately 15% more per hour than women without breast cancer, while there is no statistically significant differential for those diagnosed within the past 2 years. 5.5. Earnings In a comparison of mean self-reported annual earnings, working women who had been diagnosed with breast cancer earn on average nearly US$ 2350 more than other women, an advantage of approximately 12% (Table 1). In columns 1 and 2 of Table 4, the estimates indicate that breast cancer survivors earn more than the non-cancer control group when we estimate a log earnings equation (conditional on working). Breast cancer survivors have 21% (P < 0.10) higher earnings than the non-cancer control group. When we further controlled for years since diagnosis, women diagnosed 3 or more years prior to the interview earn approximately 26% more (P < 0.10) than the non-cancer control group. The remaining four columns of the table report estimates of similar specifications with the social security earnings data for the same year, to provide a baseline for the longitudinal analyses discussed later. Paralleling the results with the HRS earnings data, the social security earnings results suggest that working breast cancer survivors earn 39% more than the non-cancer control group. The differential is slightly smaller when we condition on positive earnings in the year prior to diagnosis (or a similar matched year for the control group). In these data, the difference between those diagnosed 2 or fewer years ago and those diagnosed 3 or more years ago is not apparent, so in the ensuing longitudinal analysis we drop this distinction. 5.6. Selection bias assessment with longitudinal social security earnings The cross-sectional results indicate that while breast cancer survivors have lower employment probabilities, those who have survived 3 or more years work more hours, earn higher wages, and have higher earnings, if they are employed. These results are surprising, if one believes that there are any long-lasting debilitating effects of breast cancer. Even if there are not long-lasting effects, the most reasonable expectation might be to find no differentials between the survivors and the non-cancer control group. On the other hand, a positive causal effect is not inconceivable, as some survivors perhaps replenish financial as16 This is based on the suggestion of Halvorsen and Palmquist (1980) for interpreting dummy variables in semilogarithmic equations. In particular, the coefficient of .12 in column 3 implies that wages are higher by {(e.12 − 1) × 100}, or 13%. This calculation is used throughout the paper in describing the results from semilogarithmic specifications.
770
C.J. Bradley et al. / Journal of Health Economics 21 (2002) 757–779
C.J. Bradley et al. / Journal of Health Economics 21 (2002) 757–779
771
sets that may have been depleted while fighting the cancer. We assess the available evidence on whether the estimated relationships described in the previous sections are causal, or are instead driven by selection bias. Selection bias could arise if women who survive breast cancer have more human capital or greater financial resources—differences that may not be fully captured in our control variables—and the unmeasured differences are associated with higher hours, wages, and earnings.17 Lower employment probabilities for breast cancer survivors appear to be more consistent with the hypothesis that we are observing causal effects. Of course the population of survivors could be heterogeneous, with some debilitated by the disease and therefore leaving the labor market, while those who remain are a select group of those most committed to the labor market, or most “robust”. Thus, additional evidence assessing the validity of a causal interpretation of the estimated effects for employment, hours, earnings, and wages is useful. We use social security earnings records that can be linked to the HRS to explore whether the cross-sectional association between breast cancer and earnings—indicating a strong positive impact—appears to be causal. In particular, we estimate three closely related longitudinal specifications to examine the association between breast cancer and earnings, preand post-diagnosis. First, we estimate the association between breast cancer (as a dummy variable) and the log of earnings in the year prior to diagnosis to see if “future breast cancer survivors” were simply higher earners even prior to diagnosis. To obtain comparable data for the control group—for use in all of the longitudinal analyses—we randomly assign an early year to each observation in this group, so as to match the distribution of the year prior to diagnosis for the treatment group of breast cancer survivors. As reported in column 1, Panel 1, of Table 5, we do not find an association between prior earnings and subsequent breast cancer. This indicates that while breast cancer survivors have higher earnings ex post (according to the estimates in Table 4), they do not have higher earnings ex ante, undermining the hypothesis that breast cancer survivors are simply a select group of high earners. For the second specification, we estimate the effect of breast cancer on the log of 1991 social security earnings, controlling for the log of earnings in the year prior to diagnosis or a matched early year for the control group (Table 5, column 2, Panel 2). In an alternative specification (Panel 2), we also control for the number of years between 1991 and the year before diagnosis, with no difference in the results. Breast cancer continues to have a positive and significant effect on earnings that is of the same magnitude (39%, P < 0.01) as in the cross-sectional analysis of social security earnings. For the third specification, reported in column 3 of Table 5, instead of using log earnings as the dependent variable, we use the difference between log 1991 earnings and log earnings in year prior to diagnosis or the matched early year for the control group. Alternative specifications were also estimated including a control for the number of years between earnings observations (Panel 2), and a control for the change in log earnings 2 and 3 years before diagnosis (Panel 3), to allow for different growth rates, and not just different levels, between breast cancer survivors and the control group. Once again our findings are robust; breast can17 The theory of compensating differentials might suggest that women in jobs with higher risk of cancer earn higher wages, although we do not think that this is a major part of the story.
772
C.J. Bradley et al. / Journal of Health Economics 21 (2002) 757–779
Table 5 Longitudinal specifications using social security earnings records Independent variables
Panel 1 Breast cancer (yes/no) Log earnings, year before diagnosis Panel 2 Breast cancer (yes/no) Log earnings, year before diagnosis Number of years between year before diagnosis and 1991 Panel 3 Breast cancer (yes/no) (Log earnings 2 years before diagnosis)− (Log earnings 3 years before diagnosis) Number of years between year before diagnosis and 1991
Log earnings year before diagnosis
Log 1991 earnings
(Log 1991 earnings)− (Log earnings of year before diagnosis)
−0.01 (0.17) –
0.33∗∗∗ (0.11) 0.47∗∗∗ (0.03)
0.33∗ (0.18) –
– – –
0.34∗∗∗ (0.11) 0.48∗∗∗ (0.03) 0.02∗∗∗ (0.004)
0.34∗ (0.18) – 0.03∗∗∗ (0.01)
– –
– –
–
–
0.35∗ (0.18) −0.03∗ (0.01) 0.03∗∗∗ (0.01)
Notes: N = 2432. Sample is restricted to women with positive earnings in “year before diagnosis” and 1991. For the control group, “year before diagnosis” is an observation from an earlier year drawn from the distribution of years prior to diagnosis for breast cancer survivors. Standard errors are in parentheses. Control variables are the same as those reported in Table 4. While earlier observations on these variables are not available, most are relatively fixed over the later part of the life cycle, although they may still enter the differenced specifications in column 3 if they are associated with differences in earnings growth. ∗ Significant at P < 0.10. ∗∗∗ P < 0.01.
cer survivors have significantly greater growth in earnings between the year prior to breast cancer diagnosis and 1991 than do women in the control group over a comparable period. Thus, the longitudinal evidence supports the conclusion that breast cancer survivorship is associated with an increase in earnings for those women who continue to work. 5.7. Further assessment of selection bias Finally, we turn to some additional evidence assessing a causal interpretation of our broader set of findings regarding breast cancer survival and labor market outcomes. First, we report analyses that compare our findings reported earlier for the probability of work, and for hours, wages, and earnings, to estimates of models where we do not control for wealth, race, age, and education. These variables have been identified as associated with surviving breast cancer and with labor market outcomes. If the exclusion of these controls does not substantively influence the estimated effects of breast cancer survival, then the argument for selection into the treatment group on unobservables is weakened. The results from this analysis are shown in Table 6. By comparing these estimates to those in the earlier tables, the reader can see that the estimated effects on employment, hours, wages, and earnings are virtually unchanged in this table.18 In contrast, if we were “worsening” the 18
The conclusion regarding earnings was also true if we used social security earnings as the dependent variable.
C.J. Bradley et al. / Journal of Health Economics 21 (2002) 757–779
773
774
C.J. Bradley et al. / Journal of Health Economics 21 (2002) 757–779
C.J. Bradley et al. / Journal of Health Economics 21 (2002) 757–779
775
776
C.J. Bradley et al. / Journal of Health Economics 21 (2002) 757–779
selection on unobservables in this table, we would expect the effects to be larger in Table 6. This provides suggestive evidence that unobservable characteristics associated with breast cancer are not influencing the labor market outcomes we observe, thus providing further support for a causal interpretation of our findings. Our next set of analyses pursue more directly the hypothesis that, among breast cancer survivors, those who exhibit higher hours of work, earnings, and wages ex post were simply more attached to the labor market and were higher-wage workers ex ante. The HRS includes job tenure information if the respondent is currently working. If the respondent is not working, job tenure information is collected for the respondent’s most recent job, unless the respondent never worked for more than a few months or last worked prior to 1972. Thus, there is a sample for which we can include job tenure as a proxy for labor market commitment or attachment. This analysis is reported in Table 7. Because there are some observations for which tenure is unavailable, we first report estimates of the earlier models using the smaller sample for which the tenure information is available (in the upper panel). Otherwise, the estimated specifications including tenure would not be comparable. Then, in the lower panel, we re-estimate the models including the tenure variable. The table reveals that the estimated differentials associated with breast cancer are largely unaffected by the inclusion of the tenure control, although they fall slightly. In our view, the fact that these relationships are little changed by including a variable strongly related to past labor market commitment also helps to undermine a selection explanation of our findings. Our final analysis tests a specific version of a selection story, namely the hypothesis that the hours differential we find arises because women working part-time are not as attached to the labor force as full-time workers. When these women are presented with a breast cancer diagnosis, these women may quit their jobs, whereas women working full-time with a breast cancer diagnosis continue to work. This could also explain the higher earnings of breast cancer survivors, as well as their higher wages (since full-time workers earn a wage premium). If this is true, we should find that breast cancer survivors are relatively more likely to work full-time, and that the hours differential should disappear when we subdivide the sample into full- and part-time workers. The results of this analysis are shown in Table 8. Columns 1 and 2 show no statistically significant differential in the probability of being employed 35 or more hours per week associated with breast cancer. In columns 3 and 4, we do not find that breast cancer has an effect on the number of hours worked per week for women employed full-time. Furthermore, in columns 5 and 6 we show a positive effect for women working less than 35 hours per week and diagnosed with breast cancer. Thus, the data are not consistent with the hypothesis that women employed part-time tend to leave their jobs after a breast cancer diagnosis. Rather, among survivors part-time women appear to work more hours.
6. Discussion Our finding that breast cancer has a negative effect on working is consistent with what one would expect given the disease and population affected, and our sensitivity analysis indicates that the coefficients on the breast cancer variables for this labor market outcome are relatively robust. However, breast cancer’s positive effect on hours worked, wages, and
C.J. Bradley et al. / Journal of Health Economics 21 (2002) 757–779
777
earnings is inconsistent with standard views of the natural course of the disease (i.e. that it should either reduce the number of hours worked and productivity because of recurrences or health effects, or at best be neutral if the cancer remains in remission and has no lasting debilitating effects), which is why we have delved at length into the question of selection bias. We believe that the various analyses of this issue are complementary, and the evidence generally suggests that our findings are not driven by selection bias. We might think that the best-case scenario is that the morbidity associated with breast cancer is negligible for at least some women, so that for these women, who are more likely to continue to work, there are no adverse consequences in terms of labor market outcomes or performance. There are several studies indicating that the morbidity associated with certain types and stages of breast cancer and its treatment may not interfere with work (e.g. Ganz et al., 1996; Satariano and DeLorenze, 1996). But this explanation is only consistent with no effects, not positive effects. One potential explanation of positive effects is that women may be unwilling to reduce number of hours worked or to exit the labor force for fear of not getting another job with health insurance benefits. Relatively few part-time jobs provide health insurance benefits (Olson, 1998). However, our analysis of the interaction between spousal insurance coverage and breast cancer and their combined effect on hours worked does not support this hypothesis, although we believe this question merits further research. Even in the absence of ideal data, our analysis improves upon what is known in the existing literature. Our study, controlling for a range of correlates, provides descriptive information on differences in labor market outcomes between cancer survivors and a non-cancer control group. We also provide objective measures of individuals’ economic circumstances after a cancer diagnosis. Finally, with respect to one outcome—earnings—we are able to observe the actual longitudinal developments a cancer survivor experiences, and to compare this with a control group. We are currently collecting longitudinal data to address this issue more fully with respect to a broader array of labor market outcomes and a wider set of cancers, and so our future research should lend further insight into the causal effects of cancer on survivors’ labor market experiences. The American Cancer Society estimates that there will be 175,000 new cases of invasive breast cancer this year among women in the United States. In general, the high incidence of certain cancer types among working age people leads to greater productivity costs for cancer compared to other categories of disease. An understanding of the labor market implications of a cancer diagnosis and treatment is particularly relevant as cancer screening (including prostate, colon, and lung, in addition to breast) is routinely applied to a working age population where these cancers may have otherwise gone undetected for some time.
Acknowledgements We are grateful to Charles W. Given, Margaret Holmes-Rovner, Daiji Kawaguchi, Donna Edwards Neumark, and two anonymous reviewers for helpful comments. Bradley’s research was supported by an NCI grant, RO1-CA86045-01A1. Neumark’s research was partially supported by NIA grant K01-AG00589.
778
C.J. Bradley et al. / Journal of Health Economics 21 (2002) 757–779
References Adams, S.J., 2001. Employer-provided health insurance and job change: new evidence of job lock. University of Michigan, Ann Arbor, Michigan, unpublished. American Cancer Society, 1999. Cancer Facts and Figures. American Cancer Society, Atlanta, GA. American Cancer Society, 2001. Cancer Facts and Figures. American Cancer Society, Atlanta, GA. Benham, L., Benham, A., 1982. Employment, earnings, and psychiatric disorders. In: Fuchs, V. (Ed.), Economic Aspects of Health. University of Chicago Press, Chicago, Illinois, pp. 203–220. Berkman, B.J., Sampson, S.E., 1993. Psychosocial effects of cancer economics on patients and their families. Cancer 72 (Suppl.), 2846–2849. Berry, D., 1993. Return-to-work experiences of people with cancer. Oncology Nursing Forum 20 (6), 905–911. Bradley, C.J., Bednarek, H.L., Neumark, D. Breast cancer and women’s labor supply. Health Services Research, in press. Buchmueller, T.C., Valletta, R.G., 1999. The effect of health insurance on married female labor supply. Journal of Human Resources 34, 42–70. Carter, B., 1994. Surviving breast cancer: a problematic work re-entry. Cancer Practice 2 (2), 135–140. Currie, J., Madrian, B.C., 1999. Health, health insurance and the labor market. In: Ashenfelter, O., Card, D., (Eds.), Handbook of Labor Economics, Vol. 3. Elsevier, Amsterdam, pp. 3309–3416. Ettner, S.L., Frank, R.G., Kessler, R., 1997. The impact of psychiatric disorders on labor market outcomes. Industrial and Labor Relations Review 51 (1), 64–81. Famulari, M., 1992. The effects of a disability on labor market performance: the case of epilepsy. Southern Economic Journal 58 (4), 1072–1087. Fow, N.R., 1996. Cancer rehabilitation: an investment in survivorship. As more people survive the disease, focus shifts on improving quality of life. Rehabilitation Management 9 (3), 48–53. Friedland, R.B., 1996. People with disabilities: access to health care and related benefits. In: Mashaw, J.L. (Ed.), Disability, Work and Cash Benefits. W.E. Upjohn Institute for Employment Research, Kalamazoo, Michigan, pp. 357–388. Ganz, P.A., Coscarelli, A., Fred, C., Kahn, B., Polinsky, M.L., Petersen, L., 1996. Breast cancer survivors: psychosocial concerns and quality of life. Breast Cancer Research and Treatment 38 (2), 183–199. Greenwald, H.P., Dirks, S.J., Borgatta, E.F., McCorkle, R., Nevitt, M.C., Yelin, E.H., 1989. Work disability among cancer patients. Social Science and Medicine 29 (11), 1253–1259. Grossman, M., 1972. On the concept of health capital and the demand for health. Journal of Political Economy 80 (2), 223–255. Gruber, J., Madrian, B.C., 1994. Health insurance and job mobility: the effects of public policy on job-lock. Industrial and Labor Relations Review 48 (1), 86–102. Halvorsen, R., Palmquist, R., 1980. The interpretation of dummy variables in semilogarithmic equations. American Economic Review 70 (3), 474–475. Heckman, J.J., 1979. Shadow prices, market wages, and labor supply. Econometrica 42 (4), 679–694. Kapur, K., 1998. The impact of health on job mobility: a measure of job-lock. Industrial and Labor Relations Review 51 (2), 282–298. Kornblith, A.B., 1998. Psychosocial adaptation of cancer survivors. In: Holland, J.C. (Ed.), Psycho-Oncology. Oxford University Press, New York, pp. 223–241. Madrian, B.C., 1994. Employment-based health insurance and job mobility: is there evidence of job lock? Quarterly Journal of Economics 109 (1), 27–54. Moffitt, R., Wolfe, B., 1992. The effect of the Medicaid program on welfare participation and labor supply. The Review of Economics and Statistics 74 (4), 615–626. Monheit, A.C., Cooper, P.F., 1994. Health insurance and job mobility: theory and evidence. Industrial and Labor Relations Review 48 (1), 68–85. Olsen, O., Gotzsche, P., 2001. Cochrane review on screening for breast cancer with mammography. Lancet 358 (9290), 1284–1285. Olsen, R.J., 1980. A least squares correction for selectivity bias. Econometrica 48 (7), 1815–1820. Olson, C., 1998. A comparison of parametric and semiparametric estimates of the effect of spousal health insurance coverage on weekly hours worked by wives. Journal of Applied Econometrics 14, 543–565.
C.J. Bradley et al. / Journal of Health Economics 21 (2002) 757–779
779
Quigley, K.M., 1989. The adult cancer survivor: psychosocial consequences of cure. Seminar in Oncology Nursing 5, 63–69. Razavi, D., Delvaux, N., Bredart, A., Autier, P., Bron, D., Debusscher, L., Stryckmans, P., 1993. Professional rehabilitation of lymphoma patients: a study of psychosocial factors associated with return to work. Support Care Cancer 1, 276–278. Satariano, W.A., DeLorenze, G.N., 1996. The likelihood of returning to work after breast cancer. Public Health Reports 111 (3), 236–241. Schagen, S.B., Frits, S.A.M., van Dam, M., Muller, J., Boogerd, W., Lindeboom, J., Bruning, P.F., 1999. Cognitive deficits after postoperative adjuvant chemotherapy for breast carcinoma. Cancer 85 (3), 640–650.